# Felo API Platform

> One API key for currently public model APIs and ready-made agent harness APIs, with current catalog information and documentation.

## Core Pages
- [Home](https://openapi.felo.ai/): Felo API Platform overview for model APIs, agent harness APIs, and pricing.
- [Models](https://openapi.felo.ai/models): Browse the currently public model APIs, current pricing, specifications and model-specific integration details.
- [Free LLM API](https://openapi.felo.ai/free-llm-api): Every model currently callable at $0, including live trials, refreshed from the catalog on each request.
- [Video models](https://openapi.felo.ai/video-models): Every video generation model currently public through Felo: live per-second rates, resolutions, clip length and the input modes each model accepts.
- [Image models](https://openapi.felo.ai/image-models): Every image generation model currently public through Felo: live per-image rates and the input and output modalities of each model.
- [Decision models](https://openapi.felo.ai/decision-models): Structured decision models that return typed answers instead of free-form text, with live token pricing and the POST /api/v1/decisions endpoint.
- [SEO API](https://openapi.felo.ai/seo-api): SEO data behind DataForSEO, Semrush, Ahrefs and Similarweb on one Felo key: SERP results, keywords, backlinks and traffic, billed per call.
- [Compare models](https://openapi.felo.ai/compare): Pick any two models and compare what you pay against list price, context and output limits, modalities and the endpoint that serves each one, read live from the catalog.
- [Felo vs OpenRouter](https://openapi.felo.ai/compare/openrouter): How Felo API Platform compares with OpenRouter on model rates, data APIs and calling conventions, with the catalog counts computed per request.
- [Felo vs LiteLLM](https://openapi.felo.ai/compare/litellm): How a self-hosted LiteLLM proxy compares with Felo as a hosted gateway: provider list price against Felo rates, what you stop operating, and how to keep LiteLLM in front of Felo.
- [Felo vs OpenAI](https://openapi.felo.ai/compare/openai): Switching from the OpenAI API to Felo: the same models priced against OpenAI's own list price, what stays unchanged, and the two lines to edit.
- [LLM gateway](https://openapi.felo.ai/llm-gateway): What the gateway is made of: three base URLs on one key, six model endpoints, model-kind routing, per-token billing, and the error and retry rules.
- [Harness](https://openapi.felo.ai/harness): Explore ready-made agent capabilities including Search, PPT, Web Fetch, X Search, Mindmap, LiveDoc, YouTube subtitles, Landing Page, and SuperAgent APIs.
- [Use cases](https://openapi.felo.ai/use-cases): Task-oriented recipes with copy-paste code: sourced research briefs, slide decks, daily digests, support knowledge bases, video-to-article, and model cost routing. Available in 26 languages.
- [Docs](https://openapi.felo.ai/docs.md): Developer documentation for authentication, AI tool setup, LLM APIs, and harness APIs. Plain Markdown index of every docs page.
- [Changelog](https://openapi.felo.ai/changelog.md): Product updates: new models, documentation, pricing, and platform capabilities. Markdown, all 26 languages.
- [APIs](https://openapi.felo.ai/apis): Directory of every third-party API available through the same key, with an endpoint list for each.
- [Data API reference](https://openapi.felo.ai/docs/data-api.md): Browse Data API categories, endpoint documentation and request examples. Markdown version generated from the OpenAPI spec.
- [Model API reference](https://openapi.felo.ai/docs/model-api.md): Model and video endpoints with protocol documentation and request examples. Markdown version generated from the OpenAPI spec.
- [Seedance video](https://openapi.felo.ai/seedance): Browse currently public Seedance routes, token rates and real video examples.
- [Pricing](https://openapi.felo.ai/pricing): USD-denominated service credits, top-ups from $1, and usage-based model and API rates. No subscription required.

## Model APIs
- [LFM3-1.7B API](https://openapi.felo.ai/models/liquid/lfm-3-1.7b): lfm-3-1.7b; LiquidAI; LFM3-1.7B is a compact 1.73-billion-parameter language model from Liquid AI, released on October 27, 2026. It uses a tri-hybrid architecture with 24 layers that mix 13 gated short convolutions, 5 sliding-window attention layers, and 6 global attention layers. Its vocabulary covers 128,000 tokens across 16 languages. The model was trained on 22 trillion tokens, a tokens-to-parameters ratio of approximately 12,717 times. It ships in six formats, including GGUF, MLX, ONNX, BF16, FP8, and NVFP4, and at 4-bit quantization it takes up roughly 1GB of memory. Reasoning is enabled by default but can be disabled via enable_thinking for direct answers.

The model is designed for local agents on phones and laptops without a dedicated GPU. It supports structured tool calling and a deferred tool search mechanism for large tool catalogs, and it has been post-trained for agentic harnesses such as Hermes Agent, Pi, and OpenClaw. In agentic evaluations, it used 58% fewer tokens per task and became stuck in fewer than 1% of attempts, versus 15% for a competing model. It outperforms the previous 2.6B-parameter LFM2.5 model while using 35% fewer parameters.

LFM3-1.7B is best suited to tool calling with structured outputs, retrieval-grounded question answering, classifying customer requests, extracting document information, and routing cases. It can be fine-tuned quickly on a single GPU for domain-specific assistants built on a company's own documents and formats. It is explicitly not recommended for agentic coding or complex multi-step reasoning, and it cannot hold as much world knowledge as larger systems, so knowledge-heavy questions should be paired with retrieval or search.; 262.144K context
- [LFM3-1.7B (free) API](https://openapi.felo.ai/models/liquid/lfm-3-1.7b:free): lfm-3-1.7b:free; LiquidAI; LFM3-1.7B is an open-weight, text-only language model from Liquid AI with about 1.73 billion parameters and 24 layers. It uses a new tri-hybrid architecture combining 13 gated short convolution layers, 5 sliding window attention layers, and 6 global attention layers. The model accepts text input and produces text output, supports tool calling and structured outputs, and handles a context of 256K tokens covering input and output combined. It supports 16 languages: English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, and Polish.

The model is designed for local agents and domain-specific applications and is post-trained for agent harnesses such as Hermes Agent, Pi, and OpenClaw. It runs on Transformers, vLLM, SGLang, llama.cpp, Ollama, LM Studio, MLX, and ONNX Runtime, and is available in BF16, FP8, NVFP4, GGUF, MLX, and ONNX quantization formats. In 4-bit quantized form it uses about 1 GB of memory, and at 128K-token context it uses about 1.6 GB of cache.

LFM3-1.7B suits local agents, tool calling, structured outputs, structured JSON extraction, answering from retrieved documents, domain-specific fine-tuning, tool selection in a 108-tool phone app, and learning new languages. Retrieval or retrieval-augmented generation is recommended for knowledge-intensive tasks. It is not recommended for agentic coding or complex multi-step reasoning.; 262.144K context
- [Decisions Fast API](https://openapi.felo.ai/models/parallel/decisions-fast): decisions-fast; Parallel; Parallel: Decisions Fast is a decisions engine from Parallel that answers named questions with typed, probabilistic outputs rather than conversational text. It combines decision models with web search, so each response is grounded in web results and comes with citations. A single request can carry multiple questions about the same content, returning probabilities for every option alongside the selected answer. The variant is built for fast, structured choices where a short, decisive result matters more than a lengthy reply.

The model accepts text, JSON, and image input and returns typed answers as output. This makes it suited to operational decisions such as lead routing, KYB rechecks, CRM field refresh, ticket routing, content filtering, deciding an agent's next step, scoring severity, and reranking candidates — tasks where a clear choice plus probability per option is more useful than prose.

Because answers are typed and probabilistic, downstream systems can act on them directly, and citations let teams trace the basis of each decision. The revision 2 variant is documented here; search results contain no information specific to it.; 32K context
- [Decisions Fast (free) API](https://openapi.felo.ai/models/parallel/decisions-fast:free): decisions-fast:free; Parallel; Parallel: Decisions Fast (free) is a text-in, text-out decisions engine from Parallel. It answers named questions with typed, bounded responses in a single request, and it can batch several questions in one API call using a shared context. Every answer includes citations, and each answer carries a confidence score. When confidence is low, the model can route uncertain cases to review. It combines the speed of System One decision models with web search so answers are grounded in live results, and it is described as the first web-grounded decisions engine, a hybrid between a search engine and a decision model.

This model is built for tasks where speed and cost matter. Typical uses include data enrichment, such as filling CRM columns with multiple questions about companies, and decision routing in workflows that classify requests into categories. It also suits any scenario that needs typed, bounded outputs backed by citations. Because several questions can be answered in one call for the cost of a single request, batch-style enrichment workloads are a natural fit.

There are limits to keep in mind. Decision models do not justify their answers, and they do not respond with "I don't know". For some question types, the confidence score reflects the probability gap between the first and second options rather than the probability of being correct. Results remain auditable through citations, and in an illustrative example a single request with shared company context answered four questions in 681 milliseconds while retrieving from 10 sources.; 32K context
- [Gemini Omni 1.2 Flash API](https://openapi.felo.ai/models/google/gemini-omni-1.2-flash): gemini-omni-1.2-flash; Google; Google: Gemini Omni 1.2 Flash is a multimodal video model from Google that takes combinations of text, images, audio, and video as input and produces video as output. It generates video from text and images, and it can edit existing video through natural-language conversation, so creators can refine a clip iteratively instead of restarting the entire creative process. Reference-driven creation lets you supply images of characters, scenes, or drawings to steer the result.

The model shows an improved intuitive understanding of physics, including gravity, kinetic energy, and fluid dynamics, and it draws on Gemini's knowledge of history, science, and cultural context. That combination makes it suited to social ad drafts, product launch clips, e-commerce motion ads, app demo videos, seasonal campaign variants, performance creative testing, short influencer-style mockups, trailers and gameplay previews, and explainers built from short prompts.

Gemini Omni Flash launched in public preview and is available in Google AI Studio and the Gemini API. At launch it generated 10-second clips, and the API initially did not support audio references or scene extension. Character consistency was limited in some scene changes and panning movements. On Video Arena it holds first place overall with an Elo of 1404, a 101 point gap ahead of Seedance 2.0 Mini.; 131.072K context
- [GPT Image 2.6 Sunburst API](https://openapi.felo.ai/models/openai/gpt-image-2.6-sunburst): gpt-image-2.6-sunburst; OpenAI; GPT Image 2.6 Sunburst is OpenAI's high-quality, precision-oriented tier of the GPT Image 2.6 series, released in 2026 and positioned alongside Flare. It accepts image input and produces image output through the Images API, and it is also available in ChatGPT and Codex across all subscription levels on desktop, mobile, and web. The model supports an image markup feature that lets users specify exact areas to change, and it maintains character, product, and brand details across edits.

The model exposes a complexity parameter that adjusts how much planning it does to satisfy a request, plus quality settings selectable from low to high. Higher-quality generation is produced at the medium or high quality settings, and SVG export makes designs directly usable on websites, apps, and print. Generation speed is improved by 50% over the previous model, and image consistency between generation and editing is stronger.

Sunburst suits detailed work where editing accuracy matters more than generation speed: animating brand mascots and characters, building GIFs that turn product shots into engaging loops, precision editing while preserving character, product, and brand details, professional visual production such as posters and product imagery, and creative workflows with multiple editing rounds. It is also available via the API as GPT-Image-2.6 Sunburst and has been added to Figma Weave.
- [GPT Image 2.6 Flare API](https://openapi.felo.ai/models/openai/gpt-image-2.6-flare): gpt-image-2.6-flare; OpenAI; GPT Image 2.6 Flare is the speed-oriented tier of OpenAI's GPT Image series, built for everyday image generation and editing. It accepts text and image input to produce images, takes up to 16 reference images for editing, and supports prompts up to 20,000 characters. Output can be rendered at 1K, 2K, or 4K resolution with a long edge up to 3840 pixels, and native transparent background generation is available. Revision 2 improves editing consistency, detail preservation, and responsiveness to sketches and image annotations, and it accepts marked-up images via the API so edit regions can be specified precisely.

The model offers six quality settings — auto, low, medium, high, xhigh, and max — and chains outputs as reference images across multi-turn editing sessions, which helps keep scenes consistent over multiple revision rounds. It applies only instructed changes while preserving subject, composition, and background, and supports natural lighting, rich textures, and complex layouts. This makes it a good fit for iterative design work, character creation with brand consistency, image fusion from multiple references, and screen or layout mockups.

Speed is a central strength: latency is reported up to 50% lower than GPT Image 2, with generation reported up to 7x faster for 1024x1080 images, and roughly 6 seconds at low quality. It is well suited to interactive and high-volume work. Known limitations include noticeable noise artifacts in some areas, and GPT Image 2 is reported to handle dense multilingual text and extreme aspect ratios better.
- [ricursiveAI/Ripple-65B-Preview API](https://openapi.felo.ai/models/ricursiveai/ripple-65b-preview): ripple-65b-preview; Ricursiveai; Ripple-65B Preview is the first model from Ricursive Intelligence, released on November 5, 2026 as open weights under the Apache 2.0 license. It is a 65-billion parameter dense Transformer pretrained from scratch on approximately 12 trillion tokens of mostly English text and code. The release includes three checkpoints: an instruction-tuned BF16 model, an NVFP4 version that quantizes MLP layers while keeping attention in BF16 with weights around 34 GiB at 4-bit, and a BF16 base model. The dense architecture was chosen over Mixture-of-Experts to reduce infrastructure complexity and enables running on consumer hardware such as a MacBook or NVIDIA DGX Spark; the BF16 model reaches 250 tokens per second on 2x B200 GPUs at batch size 8 before quantization.

The model accepts text input and produces text output, with a 1 million token context window. It supports native tool calling with one function call per turn, returning function names with JSON arguments, and a reasoning mode that returns internal reasoning as a separate section before the final answer. Optional system instructions, user and session identifiers, function definitions, sampling controls, stop strings, a numeric seed, and a toggle between deterministic decoding and sampling are all available. It handles code review and refactoring plans for files and modules up to about 15,000 tokens.

Ripple-65B Preview fits coding agents that plan changes, call tools, and continue from results, as well as code review and refactoring assistance, technical question answering, assistants that look up data through user-defined functions, and document summarization and cross-checking. Its preview release has undergone relatively light posttraining focused on agentic coding, it supports only one function call per turn, and it trails larger Mixture-of-Experts rivals on SciCode, DeepSWE, and MCP Atlas.; 1.049M context
- [Hy4 API](https://openapi.felo.ai/models/tencent/hy4): hy4; Tencent; Tencent: Hy4 is a flagship open-weight model from Tencent Hunyuan, released on August 28, 2026 under the Apache 2.0 license. It is a Mixture-of-Experts architecture with 770B total parameters and 49B activated parameters per token, built from 78 Transformer layers — the first a standard feed-forward layer and the remaining 77 MoE layers, each with 256 routed experts and 1 shared expert, activating top-8 routed experts plus the shared expert per token. Hidden size is 6144, feed-forward intermediate size 18432, and MoE intermediate size 2048, with a vocabulary of 120832 tokens. It accepts text input and returns text output, supports up to 1M tokens of context with a 64K-token output limit.

The architecture adds Gated DeepSeek Sparse Attention with IndexCache for cross-layer sparse index reuse, identity hyper-connections that expand residual paths to 4 residual streams, and a built-in native multi-token prediction layer (10B total, 0.7B activated parameters) for speculative decoding. Two response modes are offered: a high-effort reasoning mode and a no_think direct response mode. Deployment recipes for vLLM and SGLang are provided with 8-way tensor parallelism and an FP8 variant.

Hy4 is positioned for real productivity work and was tuned on data drawn from Tencent internal team workflows. It suits software engineering, office and document generation, game prototype development, and scientific research. Tencent describes it as an early, imperfect version with room for improvement in pre-training and post-training, noting known issues such as overly long reasoning chains and excessive self-verification. The API model ID is hy4-preview, accessible via Tencent Cloud International TokenHub and OpenRouter, and the model is integrated into CodeBuddy, WorkBuddy and Yuanbao. Weights are available on Hugging Face, ModelScope, GitCode and CNB.; 1.049M context
- [Reflex API](https://openapi.felo.ai/models/morph/reflex): reflex; Morph; Morph: Reflex is a small, very fast specialist model built by Morph to sit alongside larger coding models inside an agent. Instead of producing free text, it reads the agent's state once and answers every question from that single read, labeling each turn in the loop. Because its output is constrained to predefined options, answers cannot fall outside the choices provided.

The model is designed for classification tasks inside agentic workflows: labeling or categorizing turns, detecting when an agent loops, detecting when a user gives up, and detecting when a jailbreak attempt slips past. Its answers are deterministic, which makes the behavior predictable for routing and monitoring logic.

Reflex fits workflows where a quick, constrained decision is more useful than a written response. If no existing reflex covers a particular need, Morph lets you create one.; 65.536K context
- [Crouton Alpha API](https://openapi.felo.ai/models/stealth/crouton-alpha): crouton-alpha; Stealth; Crouton Alpha is a preview model available through the OpenRouter API, an OpenAI-compatible endpoint that provides access to over 400 AI models from more than 50 providers. It accepts text, image, video, audio, and file inputs and returns text output. Its context window is 1,048,576 tokens, with an output limit of 131,000 tokens, making it suited to very large inputs such as entire code repositories, design documents, call recordings, or PDFs of contracts.

The model supports tool use, document analysis, and the execution of agentic sub-tasks. It is also positioned for high-volume, repeated work made up of many small calls. Coding assistance is a further intended use. During its preview period it is offered at zero dollars per million input and output tokens.

Some caveats apply. Users report that the model can stall mid-turn during multi-step agent tasks, stopping after a short output or a few tool calls without continuing to the next steps. The provider has not disclosed the parameter count, quantization is listed as unknown, and there are no official published benchmark scores; intelligence and coding indexes are listed as not yet scored on OpenRouter. Prompts and completions may be retained by the provider but are not used for training, and EU in-region routing is unavailable.; 1.049M context
- [GLM 5.3 FlashX Legal API](https://openapi.felo.ai/models/z-ai/glm-5.3-flashx-legal): glm-5.3-flashx-legal; Z.ai; GLM 5.3 FlashX Legal is a Z.ai model built for agentic legal intelligence, combining document understanding, legal reasoning, and task planning. It accepts text, image, video, and file input and returns text, with thinking mode always enabled. It supports streaming output, function calling, context caching, structured output such as JSON, and specialized MCP services for external legal data and Word document tools. Its context window is approximately one million tokens, with 128K output tokens.

The model targets legal work end to end: reviewing contracts at the clause level for risks and conflicts, extracting key information from contracts and litigation materials, checking consistency across multiple documents, identifying legal relationships and issues in dispute, analyzing evidence, preparing cases, and drafting legal documents. It also supports due diligence working papers, defense strategies, initial contract revisions, and full legal document delivery workflows, as well as vision-driven UI coding from screenshots, page images, website URLs, or screen recordings, financial research and analysis such as earnings analysis and valuation modeling, and generating parametric 3D models from part photographs or CAD blueprints.

Z.ai documents about 200 tokens per second for faster streamed replies. The architecture uses three times less attention compute and a 4.4 times smaller KV cache than GLM-5.3 at one million token context. Thinking mode cannot be turned off, so latency and cost always include the deliberation process. The GLM-5.3-Flash family has 320 billion total parameters and 18 billion active parameters; FlashX carries the official API ID glm-5.3-flashx. It is open-source under the MIT License, can run locally with dynamic GGUF quantizations, with a one-bit version running on a 100 gigabyte machine and a three-bit version fitting a 128 gigabyte Mac or NVIDIA DGX Spark.; 1.049M context
- [Clairic API](https://openapi.felo.ai/models/cloudflare/clairic): clairic; Cloudflare; Clairic is an extraction model from Cloudflare, hosted on Workers AI under the model ID @cf/cloudflare/clairic. It takes text, optional images, and a typed field schema as input and returns strongly typed values in a single pass — text spans, numbers, dates, times, durations, lists, estimates, and predictions — along with a confidence score and null probability for each field. It does not generate output tokens. Cloudflare positions it as a new System One class of AI model and as the first model applying that concept to extraction tasks, inspired by the Jev model announced by TypeSafe.

The model is built for pulling structured data out of documents such as invoices, receipts, and insurance paperwork. It also supports financial data aggregation, date and duration calculation, inventory forecasting, and information extraction in customer service. Cloudflare reports higher accuracy, lower latency, and lower cost than conventional LLMs on data extraction tasks, and says it handles one-step or two-step calculations such as sums, differences, conversions, and date math reliably. In a three-stage pipeline test for insurance claim processing, it reduced workflow time from 61.7 seconds to 19.3 seconds, a 69% latency reduction, while cutting cost by about 49%; it also succeeded at estimating repair costs where Gemma 4 and Qwen pipelines failed.

Clairic is intended as a component of a broader AI architecture rather than an LLM replacement, and works alongside LLMs and the Clef decision model in a three-stage pipeline. It is best suited to well-defined extraction and simple calculation tasks, and should be combined with LLMs for complex reasoning and text generation. The preview version still struggles with multi-step arithmetic and long dependency chains or aggregation across multiple table rows. Estimates depend on comparable data, history, or descriptive text in the input and are only limitedly accurate without evidence, while prediction is basic and does not replace time-series models trained on specific data. It is meant for general pattern matching and non-critical uses. Clairic is currently available as a preview, with an open-weights release planned in the near future.; 65.536K context
- [Mischief 1 API](https://openapi.felo.ai/models/tbc/mischief-1): mischief-1; TBC; TBC: Mischief 1 is offered by The Biological Computing Company and is available through the OpenRouter API. It is aimed at early-stage creative and exploratory work rather than heavy production output.

The model suits ideation, previsualization, moodboarding and variant testing, making it useful when you want to explore several directions quickly before committing to one. According to the available information, it provides good signal for what real production workloads look like.

Usage is priced from 0.005 US dollars per second. Note that a MiSCHiEF benchmark exists as a minimal-pairs safety and culture evaluation, but it is a different entity unrelated to this model.
- [Joey 1 API](https://openapi.felo.ai/models/baseten/joey-1): joey-1; Baseten; Joey 1 is a 2B-parameter open-weight decision model from Baseten, post-trained from Qwen3.5-2B. Rather than generating free-form text, it accepts structured context as input and returns calibrated probabilities for predefined choices. It handles classification, routing, and scoring in a single forward pass, and supports synchronous API calls for real-time use. Because the weights are open, it can also be downloaded and run in-house.

This design makes it well suited to high-volume first-line decision making: content moderation pipelines, fraud detection, ticket triage and routing, spam and phishing detection, and prompt classification. A probability threshold of 0.85 is recommended when treating outputs as high-confidence decisions. It is intended for decision tasks rather than general text generation, so open-ended writing is out of scope.

Joey 1 serves as the fast, cost-efficient high-volume tier of a hierarchical setup, where uncertain cases are escalated to larger models such as Elephant-1 or Mammoth-1. It is part of Baseten's model library and deployable as an API endpoint, backed by an inference stack with custom kernels, modern decoding techniques, and advanced caching, 99.99% uptime, fast cold starts, and scaling in any region and any cloud, including hybrid setups.; 16.384K context
- [Riverflow V2.6 Pro API](https://openapi.felo.ai/models/sourceful/riverflow-v2.6-pro): riverflow-v2.6-pro; Sourceful; Sourceful: Riverflow V2.6 Pro is a text-to-image and image-editing model built for commercial design work. It generates structured visuals such as packaging mockups, marketing collateral, product presentations, and design systems. The model emphasizes clean geometry and layout logic, keeping compositions spatially consistent and rendering text accurately — qualities that matter when a visual has to carry product information and brand structure rather than function as free-form artwork.

Because it handles both generation from text and editing of existing images, Riverflow V2.6 Pro fits workflows where a layout is drafted, revised, and reused across collateral. Packaging mockups and marketing assets are its clearest use cases, followed by presentation visuals and the system-level assets that keep a design language consistent.

Official details for the V2.6 Pro version specifically were not found. The Riverflow family also includes Riverflow 1, Riverflow 2.0, Riverflow 2.0 Pro, Riverflow 2.0 Fast, Riverflow 2.5, Riverflow 2.5 Pro, and Riverflow V2.5 Pro.
- [Particle-1 API](https://openapi.felo.ai/models/particle-ai/particle-1): particle-1; Particle AI; Particle AI: Particle-1 is an agentic coding model trained to handle software engineering work. It is post-trained on DeepSeek V4.1 Flash, which serves as its foundational architecture, and it is built to generate concise outputs suited to agentic coding workflows. It is Particle AI's first model.

The model targets automated code generation, code refactoring, and bug fixing. Its main selling point is token efficiency: it uses roughly 45 percent fewer tokens on software engineering tasks than other models, which lowers costs for teams running large volumes of coding tasks and typically means faster generation times. Particle AI positions it as a token-efficient, cost-effective alternative in the AI coding model market.

Because it is tuned for compact responses, Particle-1 fits best in agent loops and automated pipelines where every token counts. No public benchmark scores are available for the model, so evaluations should be based on your own coding workloads rather than published rankings.; 1.049M context
- [Kling 4.0 Image-to-Video API](https://openapi.felo.ai/models/kling/4.0-image-to-video): kling-4.0/image-to-video; Kling; Kling 4.0 Image-to-Video is Kling AI's unified creative model for turning visual references into finished video. It animates a reference image with motion, camera work, and storytelling, and generates native video clips from 3 to 30 seconds. Up to 10 keyframe images can steer key visual moments, while a single project can draw on up to 10 images and 5 videos as references — including subject images, three-view references, image grids, infographics, keyframes, and wireframes.

The model is built for control and stability. It delivers stable large-scale motion and complex camera work such as fast movement, tracking shots, and sweeping orbital movements, while preserving subject structure and spatial relationships. Characters render realistically with improved facial details, materials, and lighting, and expressions and posture carry greater nuance. Output includes two-channel stereo audio with improved lip-sync accuracy, and multilingual text, logo, and emoji generation is improved.

It suits brand advertising and commercials, short-form storytelling and cinematic scenes, product campaigns, social media posts, and concept videos or scene previews built from storyboards and visual references. Multimodal editing of subjects, backgrounds, and visual styles works through text and visual references, and an existing video can serve as a reference for pacing, shot composition, and camera movement. A unified input box combines text, images, videos, and audio references, alongside Grid and List views for comparing generations, timeline-based editing, and a Canvas and Agent experience for organizing creative projects. It is available via API and through cloud catalogs such as Atlas Cloud.
- [Kling 4.0 Text-to-Video API](https://openapi.felo.ai/models/kling/4.0-text-to-video): kling-4.0/text-to-video; Kling; Kling 4.0 Text-to-Video is Kuaishou's text-only entry point to the Kling 4.0 family. It generates a 3 to 30 second clip from a written prompt alone — no source image or reference footage required — at 720p or 1080p resolution and in four aspect ratios: 16:9, 9:16, 1:1, and 21:9. Native stereo audio is rendered in sync with the picture, including two-channel dialogue, sound effects, and ambience, with improved lip sync for characters speaking in multiple languages.

The model accepts either a single text prompt or a multi-shot storyboard, letting you script several shots with individual durations into one continuous video. Prompts of up to 26,000 characters can cover subject, motion, camera direction, and sound in one block, backed by prompt understanding of up to 8,000 tokens. A single generation task can draw on up to 15 reference assets — up to 10 images, 5 videos totaling up to 30 seconds, 7 elements, and 3 video elements — plus up to 10 keyframe inputs for defining key visual moments.

It is built for short-form and social video, advertisements and product demos, multi-shot narrative content, and previsualization. Large-scale motion stays stable through push-ins, pull-outs, pans, tracking shots, and orbit shots, and output remains coherent during high-speed action. A Flash mode offers faster generation at lower cost with videos up to 20 seconds at maximum 720p resolution, though it does not support start-and-last frames or multiple keyframes.
- [d1.1 API](https://openapi.felo.ai/models/liquid/d1.1): d1.1; LiquidAI; LiquidAI: d1.1 is a decision model built on a larger open-weight base with Liquid-specific customizations. Rather than producing prose, it answers structured questions directly: yes/no questions return a probability between 0 and 1, choice questions return a probability for each label, and score questions return an expected position across ordered levels. It can take several questions about the same state in one request and computes all option probabilities in a single forward pass with zero output tokens. It accepts text and image input and produces text output.

The model is aimed at fast structured decision-making inside software. Typical uses include classification, routing, scoring, support ticket triage, content moderation, response evaluation, image inspection, and gating agent steps. Text decisions take 200 to 300 milliseconds, which is fast enough for real-time applications, and the model is positioned as a replacement for costly calls to large language models when only a structured decision is needed. It reports equal or better performance than GPT-6.1 Sol and Claude Opus 5.5 in 4 of 6 real applications, while responding substantially faster than both on all tasks.

Other reported strengths include strong multilingual evaluation results, high resistance to prompt injection, and effective handling of longer inputs. In interactive tests, adding the screen alongside text raised a Tetris score from 70 to 81; in Wordle it read the board directly from screenshots and solved 12 of 12 games in an average of 3.8 guesses; and in Quick, Draw! it recognized 5.2 of 6 doodles from 62 words.; 65.536K context
- [GPT-6.1 Terra Pro API](https://openapi.felo.ai/models/openai/gpt-6.1-terra-pro): gpt-6.1-terra-pro; OpenAI; GPT-6.1 Terra Pro is an OpenAI model listed in the manufacturer's model catalog. The available material confirms little beyond its presence in that list: one source verified only two pricing rows from OpenAI's own page and explicitly confirmed no capability claims, no context window, no maximum output size, and no usage conditions for this Pro variant.

Because of that gap, the Pro model's strengths and intended tasks cannot be stated with confidence. The closest documented reference point is the non-Pro GPT-6.1 Terra, a mid-tier model in OpenAI's GPT-6 series released on October 26, 2026, with a 1.1M token context window. That model is described as suited to everyday agentic work, coding, and writing, with particular strength in reading screenshots and on-screen content, and is recommended as the default for high-volume document processing, form entry, and computer-use work.

Those details describe GPT-6.1 Terra, not GPT-6.1 Terra Pro, and should not be transferred to the Pro variant. Anyone evaluating GPT-6.1 Terra Pro should check OpenAI's current model documentation for verified specifications before relying on it for production workloads.; 1.05M context
- [GPT-6.1 Terra API](https://openapi.felo.ai/models/openai/gpt-6.1-terra): gpt-6.1-terra; OpenAI; GPT-6.1 Terra is a mid-tier model in OpenAI's three-tier GPT-6.1 series, positioned between the flagship GPT-6.1 Sol and the fast, low-cost GPT-6.1 Luna. It combines detailed visual reasoning with accurate computer operation and writes clearer, more natural prose. Announced and made available on October 26, 2026, through ChatGPT Work, Codex, and the API, it is designed as a fast and efficient option for everyday professional work.

The model supports text and image input with text output. It handles scoped code fixes, supports computer-use tasks with clearly defined outcomes, and is suited to agentic codebase investigation and iterative solution development. Everyday applications include drafting proposals, interpreting technical drawings, filling in forms, video editing, matching product photos with catalogs, and understanding complex documents.

At 16.1 seconds per task, it is reported as the fastest OpenAI model and the third fastest overall. Its error rate on error-inducing prompts at low reasoning effort is 16.7%, down from 53.6% for GPT-5.6 Terra. At max effort, the error rate is 4.3%. In the lineup, it sits below the flagship GPT-6.1 Sol and the top-tier GPT-6 Astra.; 1.05M context
- [Kling 4.0 Omni-Video API](https://openapi.felo.ai/models/kling/4.0-omni-video): kling-4.0/omni-video; Kling; Kling 4.0 Omni is Kuaishou's reference-driven variant of the Kling 4.0 video generation model. While base Kling 4.0 generates video from text prompts, a single image, or keyframes, Omni creates and edits video from reference assets. In a single task it can combine up to 15 multimodal references: as many as 10 images of up to 30 MB each, up to 5 video clips of 3 to 30 seconds with a combined length no longer than 30 seconds, up to 3 voice profiles, and up to 7 saved elements, 3 of which may come from video. This reference-driven approach helps maintain consistency of characters, products, and scenes across shots.

Outputs run from 3 to 30 seconds, default to 10 seconds, and adjust in 1-second steps. Resolution options are 720p, 1080p, and 4K, with aspect ratios of 21:9, 16:9, 1:1, 9:16, and auto, where auto follows the reference material's ratio. Video is 10-bit HDR with dual-channel stereo audio. Multi-shot narrative is enabled by default with automatic shot splitting, up to 10 keyframes allow precise control of key moments, and targeted local edits are supported. A creative remake feature reuses an existing video's rhythm, shot structure, and style.

The model suits professional video production that benefits from 4K, 10-bit HDR, and stereo audio. Typical tasks include brand advertising built on an existing product video's structure and rhythm, social media content following trending video structures, narrative video that needs character consistency across shots, product demo videos that combine product images and scene video, editing existing footage, and integrating new clothing and object references without changing original motion or character. Users note that Seedance may still lead in realism and prompt precision.
- [MoonshotAI: Kimi K3.1 Pro API](https://openapi.felo.ai/models/moonshotai/kimi-k3.1-pro): kimi-k3.1-pro; Moonshot AI; Kimi K3.1 Pro is the first model in Moonshot AI's K3.1 series and the post-trained successor to Kimi K3. It is a flagship open-weight multimodal reasoning model built on Kimi K3's Mixture-of-Experts design, with 2.8 trillion total parameters and 104 billion activated parameters per forward pass. The architecture incorporates Kimi Delta Attention and Attention Residuals, a Stable LatentMoE framework that activates 16 out of 896 experts, and MXFP4 weights with MXFP8 activations under quantization-aware training. A MoonViT-V2 vision encoder with 401 million parameters handles native image and video understanding alongside text. Full model weights are released under the Kimi K3 License.

The model accepts text, image, and video input and produces text output, with streaming, structured generation, and tool calling with configurable reasoning effort. It can run scripts for Blender and CAD software without MCP connectors and coordinate terminal tools. Compared with Kimi K3, Moonshot AI reports substantial gains in coding, 3D creation, financial research, and agentic capabilities, positioning the model for long-horizon engineering with minimal human supervision.

Kimi K3.1 Pro suits demanding knowledge work that spans many steps: long-horizon coding, GPU kernel optimization, compiler development, chip design, CAD and parametric modeling, vision-in-the-loop game development, financial research, and multi-step business workflow automation. It is available on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.; 1.049M context
- [Flint (free) API](https://openapi.felo.ai/models/recursive/flint-v1.0:free): flint-v1.0:free; Recursive; Flint is Recursive's first model: an open-weight mixture-of-experts language model built for coding agents. It has 39.8 billion total parameters with approximately 2.5 billion active per token, arranged across 49 layers — 2 dense layers followed by 47 mixture-of-experts layers. It accepts text input and produces text output, supports multi-turn tool use, and ships under the Apache 2.0 license, so it can be used commercially, modified, and redistributed. Recursive says Flint was largely built by its own AI research system, Eureka Machine.

The model is aimed at agentic coding workflows: investigating codebases, fixing bugs, implementing features, and writing tests. Its 262,000-token context is shared between prompt and output, enough to hold a meaningful slice of a mid-sized repository alongside the agent's reasoning trace without aggressive chunking or retrieval tricks; Recursive recommends an output budget of at least 32,768 tokens. Typical uses include helping developers write code, building small applications, and solving programming problems.

Flint can run on a single B300 or H200 GPU per replica using an SGLang serving kit, and Recursive reports throughput of 125 or more tokens per second on an M4 Pro and 150 or more on a 64GB NVIDIA GPU. The BF16 checkpoint weighs around 80 GB. It was free to access on OpenRouter at launch.; 262.144K context
- [Drex Max API](https://openapi.felo.ai/models/nace-ai/drex-max): drex-max; Nace.AI; Nace.AI: Drex Max is the flagship multimodal decision model in the Nace.AI product family. Built as a 27B dense model based on Qwen3.8-27B, it uses a non-causal full attention mechanism and 255-direction readout heads. Instead of generating free text, it answers typed questions in a single forward pass: yes/no questions return the probability of yes, choice questions with 1 to 255 options return the most likely option with per-option probabilities and confidence, and ordered-level questions with 1 to 255 levels return the expected level with per-level probabilities, a legend and confidence. It never returns answers outside the options provided.

The model accepts text, JSON and images, and each question can reference all images in the request. It is compatible with TypeSafe System One and shares the same request and response format as Jev. Long states above 8,192 tokens are truncated to the most relevant parts, and only tokens actually read are billed. Drex Max is designed for decision-making in computer and browser use agents, reading information from screens, charts, forms, photos and video, customer support urgency classification, routing inquiries to departments, and assessing customer frustration levels.

It is licensed under the Nace AI Open RAIL-M License (modified) and is available on OpenRouter. Fully automated decision-making is prohibited in employment, credit, housing, insurance, healthcare, education, immigration and law enforcement without meaningful human review, as are discrimination based on legally protected characteristics, medical, legal or financial advice and decisions without qualified human oversight, and use for military, weapons, mass surveillance, fraud or spam purposes.; 262.144K context
- [MiniMax M3.1 Flash API](https://openapi.felo.ai/models/minimax/minimax-m3.1-flash): minimax-m3.1-flash; MiniMax; MiniMax M3.1 Flash is MiniMax's next-generation Flash model, belonging to a series built for high-frequency, high-intensity, large-scale agent calls. It accepts text, image and video input natively and produces text output, with a context window of 1M tokens. Thinking is always on and adjustable across five effort levels—low, medium, high, ultra-high and max—with max as the default. A sparse attention architecture reduces the compute cost of response generation, and the model is available through OpenAI-compatible and Anthropic-compatible API endpoints with tool calling support.

The model is designed for agentic work. It can execute tasks across SaaS systems, plan across multiple steps with cross-document consistency, and handle long agentic coding sessions including repository generation, refactoring and terminal tasks. It also targets computer use in real environments, cybersecurity work such as code auditing, vulnerability patching, penetration testing and malware analysis, plus frequent coding and research tasks spanning financial research, multimodal understanding and frontend development.

In everyday use, output speed exceeds 200 tokens per second, with median output speed above 170 tokens per second. Online inference is about 3x faster than the previous generation, while overall usage cost on agentic tasks drops to roughly 30% of the previous generation. An early anonymous version, Space Bunny, processed over 200 trillion tokens. Independent testing has flagged limitations: Artificial Analysis reported an intelligence index of about 42 to 43 and a high overconfident wrong-answer rate, and one user review says benchmark results do not hold up in real-world comparison. Fact-checking is advised where factual accuracy matters.; 1.049M context
- [DeepSeek V4.1 Flash 1017 API](https://openapi.felo.ai/models/deepseek/deepseek-v4.1-flash-1017): deepseek-v4.1-flash-1017; DeepSeek; DeepSeek V4.1 Flash 1017 is a multimodal mixture-of-experts model from DeepSeek that accepts text and image input and produces text output. It supports up to one million tokens of input context and a maximum output of 384,000 tokens. Reasoning effort is continuously controllable, so users can tune how deeply the model thinks for a given task. The model is released under the MIT license, which permits commercial use and self-hosting, and it is available as open weights that can be downloaded and run on your own hardware.

The model does not rely on a Jinja-format chat template. Instead, it provides a self-contained Python reference implementation for prompt encoding, and it ships deepseek-recipe, a set of Rust libraries with Python bindings that offer the same prompt format. One source reports 763.2 billion total parameters, while another lists 552 billion backbone parameters with 8 billion active during prefill and 16 billion active during decode. The model was trained on 45 trillion tokens, and its knowledge cutoff is May 2025.

DeepSeek V4.1 Flash 1017 fits long-context multimodal reasoning, coding, document understanding and analysis, chart interpretation, visual question answering, and tool-using agentic applications that consume text and images and generate text. It is positioned as the cost-efficient tier of the V4.1 family. Reports indicate it exceeds DeepSeek V4 Pro on performance, speed, and task completion time, and that it outperforms K3, GLM-5.3, and GPT-5.6 Sol in many cases, though it trails Opus 5 on several benchmarks.; 1.049M context
- [Glimpse API](https://openapi.felo.ai/models/elorian/glimpse): glimpse; Elorian; Elorian: Glimpse is Elorian's first model, released on October 16, 2026 as a research preview with API access and a waitlist-only chat interface. It accepts text, image, and video input and produces text output, backed by a 262K-token context. Its defining trait is a visual scratch space: rather than relying on a single lossy first impression, it revisits the actual images and videos mid-reasoning, zooming into image corners to inspect details, slowing video to a higher frame rate for review, and tracking a specific object across multiple frames.

The model is built for visual workloads that demand close attention to detail. It suits synthetic data labeling and high-level planning in robotics, video review on construction and manufacturing sites, dense video captioning for media archives, accurate object counting on construction sites, timestamped video review, and comparing multiple images. It supports tool calling and structured outputs, and Elorian positions it for applications that perform frequent, large-volume visual analysis across both physical-world and media use cases. Elorian says it fills a gap where frontier general models are either insufficiently accurate or too expensive for visual workloads, and notes that Glimpse reports when it cannot identify an item.

Independent hands-on testing tempers the picture. Glimpse does not outperform GPT-6 Astra and Claude Opus 5.5 on every task; GPT-6 Astra remains stronger on fine visual puzzle tasks and dense video captioning, while Claude Opus 5.5 is slightly better at reading scientific charts. Glimpse can also contradict its own earlier statements within long multi-part answers, so it is not fully self-consistent. Elorian's own comparisons claim Glimpse surpasses GPT-6 Astra on scientific chart reasoning, open-vocabulary object counting, short video captioning, and manipulation event understanding, and matches it on embodied reasoning, though no scores are given.; 262.144K context
- [Qwen Image 2.1 Turbo API](https://openapi.felo.ai/models/qwen/qwen-image-2.1-turbo): qwen-image-2.1-turbo; Qwen; Qwen Image 2.1 Turbo is a distilled accelerated checkpoint in the Qwen-Image-2.1 series, built on the same 7B vision generation architecture as the base model. It uses knowledge distillation to compress denoising to an 8-step sampling schedule at CFG 1.0 while keeping 2K-level image generation. The model runs at BF16 native precision and is about 14.2GB in size. Weights are available on Hugging Face and ModelScope and can be loaded through the Diffusers QwenImage21Pipeline.

It supports text-to-image generation and image editing driven by natural language instructions, with up to 4 additional reference images per editing request. It also supports RGBA output for transparent-background images. For editing, it offers stronger restoration of facial details and better consistency of product text and textures across new scenes. The recommended generation resolution is about 2 megapixels, such as 2048x2048 at 1:1.

This model is best suited to high-frequency local image generation workflows and rapid iteration, including product visual design for posters, concept art, and ad assets, as well as e-commerce product imagery and model outfit changes. It must use its dedicated 8-step sampling schedule, as the standard default schedule produces different images. At 4-bit quantization and below, output deviates noticeably from the original, and higher-quality quantized text encoders are recommended because heavily quantized encoders degrade output. It is released under the Qwen Research license, which carries restrictions on commercial use.
- [Qwen Image 2.1 Pro API](https://openapi.felo.ai/models/qwen/qwen-image-2.1-pro): qwen-image-2.1-pro; Qwen; Qwen Image 2.1 Pro is a professional, closed version of Alibaba Qwen team's image model, released on September 20, 2026. It combines text-to-image generation with prompt-guided editing in a single model, and it can compose from up to 10 reference images. Prompts are rewritten by a large language model before generation, and the model supports both a direct mode and an agent mode with optional thinking. It generates natively at 2K resolution with RGBA output, so transparent backgrounds are available without post-processing. The positive prompt accepts 2 to 3000 characters, and three region-specification methods are offered for editing: colored circles drawn on the image, painted mask regions, or an original image plus a separate mask.

The model is aimed at professional visual production: product photography, marketing visuals, multi-image composition such as merging six portraits into a group photo or ten furniture photos into one room scene, transparent-background icons, game UI assets, and localized retouching. Qwen reports that it leads on background structure stability during iterative edits and recommends it for transparent image generation and locked-background editing. On Qwen's own undisclosed Qwen-Image-Bench, it scores 62.36 points, above Qwen-Image-2.1 at 60.28 and Nano Banana 2.0 at 59.82.

Those benchmark figures are self-built and undisclosed, with no table, sub-scores, or scoring method, so they are Qwen's claims rather than independent measurements. User tests report lower generation quality than GPT Image 1.0, with more errors and artifacts, and the official release does not provide per-image generation seconds, VRAM usage, throughput on specific GPUs, or API pricing details.
- [Decider Flash V1.2 API](https://openapi.felo.ai/models/perplexity/pplx-decider-flash-v1.2): pplx-decider-flash-v1.2; Perplexity; Perplexity: Decider Flash V1.2 is a compact, open-weight decision model distilled from Perplexity's larger decision models and built on the Qwen3.5-4B backbone with 4 billion parameters. Rather than generating explanatory text, it answers named yes/no questions with probabilities, choice questions with a probability per option and a most-likely selection, and rubric-score questions with a probability per level plus an expected score. It accepts text, JSON, image, and PDF input, supports structured output, and shares the same API contract as the full Decider V1.2 model. A 262K token context window is notable for a model of this size, and it processes a whole request in a single pass so multiple questions about the same input can be answered together. It was trained with model souping, which averages weights across training runs to improve robustness.

The model is intended for high-volume, decision-style workloads where software compares a numeric answer against a threshold: classification, routing support tickets to the right teams, filtering, labeling messages, content moderation, sentiment scoring, rubric grading, and structured data extraction or attribute verification from images. Because it outputs probabilities, it fits pipelines that need a confidence value rather than free-form commentary, and it can be run locally on compatible hardware at about 13 GB of VRAM at 4-bit quantization or on-device through the Decider for Mac app, which requires 5.8 GB of storage and 16 GB of memory. It is available through Perplexity's API, the Opper gateway, or locally, and is positioned as the cost-efficient option in Perplexity's decision model lineup.

Users should keep several constraints in mind. The model is designed for one question per request; multi-question requests process and bill the state again for each question. No explicit reasoning mode is documented. Training used sequences up to about 32K tokens, with longer contexts spot-checked rather than broadly validated, so real-world accuracy on very long inputs may vary. Perplexity recommends validating accuracy on your own labeled examples, especially for nuanced judgments, and independent testing of the larger 27B Decider model showed weaknesses in contextual visual judgment tasks.; 262.144K context
- [XingChen: Xing4.0 29B A4B (free) API](https://openapi.felo.ai/models/xingchen-agi/xing4.0-29b-a4b:free): xing4.0-29b-a4b:free; Xingchen AGI; Xing4.0 29B A4B is a lightweight agentic large language model released by China Telecom Artificial Intelligence Technology Co., Ltd. on September 17, 2026. It uses a Mixture-of-Experts architecture with 29B total parameters and 4B active parameters per inference, supports text input and text output, and offers a natively 256K context that can be extended to 512K. Reasoning is enabled by default and can be turned off. The model is open-sourced under the Apache-2.0 license and succeeds the TeleChat series, now renamed Xing.

The model is built for agentic work: it handles complex task understanding, task planning, tool calling, and autonomous execution. It is well suited to code development, data analysis, office automation, life services, and agent applications that need multi-step planning and independent operation. Benchmarks show 75.0 on SWE-bench Verified and 57.5 on Terminal-Bench 2.1, and it ranked third on SuperCLUE Agent with 93.52. It outperforms same-size competitors on SWE-bench Verified and Terminal-Bench 2.1.

Xing4.0 29B A4B is the first fully domestically produced lightweight agentic large model, trained entirely on Ascend compute and the MindSpore framework. It can run complex autonomous workflows on a single consumer desktop GPU, and after 4-bit quantization it needs about 15GB of VRAM for local deployment. A GGUF version runs in local tools such as LM Studio with at least 32GB of memory. It supports private deployment, is compatible with LLaMA-Factory and Mindformers fine-tuning frameworks, and is adapted to SGLang, vLLM, and KTransformers inference frameworks. Weights and configuration files are provided in Hugging Face Transformers format.; 262.144K context
- [FLUX 3 Image Klein API](https://openapi.felo.ai/models/black-forest-labs/flux-3-image-klein): flux-3-image-klein; Black Forest Labs; FLUX 3 Image Klein is an open-weight image model from Black Forest Labs, released on October 15, 2026. It accepts text and images and returns images, covering text-to-image generation, editing from a single reference image, and composition from up to 10 reference images. Bounding boxes on a 0 to 1000 grid steer the placement and scale of elements without hard-masking pixels, and the model can render text inside generated images. A prompt upsampler expands short requests into detailed scene captions and is on by default in the CLI, demo and ComfyUI; using the model without it significantly weakens results.

The image transformer has 5.6B parameters and pairs with Qwen3-VL-4B-Instruct as a separate text encoder. A distilled checkpoint runs in 16 steps without guidance and backs the BFL API, while the undistilled Klein Base checkpoint uses 50 steps at CFG 3.5 for fine-tuning, LoRA training and research. Weights are 11.2GB in BF16 and 5.6GB in FP8, and an INT8 build runs with 8GB of VRAM on NVIDIA GeForce RTX GPUs via ComfyUI or Diffusers, generating a 1-megapixel image in about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070.

It suits creators and developers who need fast local or on-device image editing, multi-reference composition, and deployment on their own infrastructure for data sovereignty. Internal evaluations place it first for text-to-image, multi-reference generation and image editing, ahead of larger open-weight models and above Ideogram V4.5 and FLUX.2 pro. Non-commercial use is permitted under the FLUX Kommunity License v1.0, with commercial output exceptions for eligible users subject to conditions and restrictions; other uses may require a separate license from Black Forest Labs.
- [Jev 1.14 API](https://openapi.felo.ai/models/typesafe/jev-1.14): jev-1.14; TypeSafe; TypeSafe: Jev 1.14 is a System One model from TypeSafe AI, released on October 14, 2026. It specializes in fast, reliable decisions rather than complex reasoning, returning typed answers that code can branch on directly instead of generating free text. The model supports four question types: noul (yes/no with probability), choice (select from defined options), score (numeric score), and box (object locations as top, bottom, left, and right fractions from 0 to 1 plus confidence). It is served via the POST /v1/decisions endpoint and offers a 262,144-token context window, expanded to 4x the previous size.

The model is suited for model-to-model routing, content classification and moderation, multilingual classification and moderation, image-based decisions such as product photo classification and screenshot flagging, and locating tables and charts in document analysis pipelines. Overall quality improved over the previous version, with the largest gains in non-English language performance. Average latency is 0.18 seconds, and it is reported to be about 5x faster and 25x cheaper than conventional LLMs. On System One tasks, it reaches intelligence comparable to existing LLMs.

Image input and long context are experimental features enabled by the TypeSafe-Danger-Zone-Images: true and TypeSafe-Danger-Zone-Long-Context: true request headers, respectively. TypeSafe states these features are not suited for production use. For box, object presence and location are separate checks; returned coordinates do not prove the object exists. Public benchmark scores are limited and no overall composite score is published. Developers are advised to set confidence thresholds from their own labeled data and keep human review and application permission checks. Fixed model IDs are recommended when probabilities feed a fixed threshold; typesafe/jev-1.14 is pinned to version 1.14, while typesafe/jev is an alias for the latest Jev release.; 262.144K context
- [Gemini 4 Argon API](https://openapi.felo.ai/models/google/gemini-4-argon): gemini-4-argon; Google; Gemini 4 Argon is Google's first frontier model in the Gemini 4 generation and sits at the top of its lineup. It accepts text, image, video, audio and PDF input and produces text output, with a context window of 1,048,576 tokens. By default it generates up to 65,536 tokens, extendable to 1,048,576 tokens with long decode. It supports thinking at low, medium and high levels, along with function calling, structured outputs, code execution, file search, search grounding and Google Maps grounding.

The model is designed for long-running complex work in software engineering, enterprise knowledge work and cyber defense. It handles coding, finance tasks such as multi-step ledger reconciliation and variance analysis, legal review of case law and regulatory filings, cybersecurity defense and remediation automation, robotics control and physical understanding, and vision tasks including object detection and OCR. Google reports top scores on 14 of 18 benchmarks in its own evaluations, and the model is used for much of Google's internal engineering work.

Gemini 4 Argon offers concise, low-repetition chat responses and frequent progress updates on long tasks. It does not support audio generation, image generation, the Live API or explicit caching. Its knowledge cutoff is March 2026, and it may exhibit limitations common to foundation models such as hallucination, as well as occasional latency and timeout issues. It was released in stages, first to trusted cyber defenders through the Fairwind Program, and participated in the US government's voluntary pre-release model access process.; 1.049M context
- [Muse Wonder API](https://openapi.felo.ai/models/meta/muse-wonder): muse-wonder; Meta; Muse Wonder is a frontier AI model released by Meta in October 2026. It accepts text, images, video, audio and PDFs as input and returns text. Its 1,048,576-token context window supports long documents and extended multi-step work. Core capabilities include Active Memory, which recalls important details from past conversations and surfaces relevant information at the right time, and Active Focus, which keeps the model working toward a goal when a task becomes difficult by trying other approaches and adapting to new information.

The model is built for agentic workflows and multi-step task handling. It asks the user for confirmation or checks in before important actions when needed. For coding, it plans long work and splits tasks among helper agents, while Muse Code handles tool execution and approvals. Meta positions Muse Wonder as a step toward personal superintelligence and as the result of a 14-month rebuild of its AI technology; the company plans to use it to enhance Muse, its personal agent product, starting with subscribers.

Muse Wonder suits prototyping, coding, writing and creative work as well as interpreting legal, financial and health information. Use cases include understanding lease agreements, legal work evaluated on Harvey's Legal Agent Benchmark, planning and executing long-horizon coding projects, and restaurant reservations and availability checks. It outputs text only and cannot generate images, video or audio, and third-party evaluations report it is somewhat weaker at writing, front-end design and creative work.; 1.049M context
- [Claude Haiku 5.5 API](https://openapi.felo.ai/models/anthropic/claude-haiku-5.5): claude-haiku-5-5; Anthropic; Claude Haiku 5.5 is the successor to Claude Haiku 4.5 and the smallest, fastest and lowest-cost member of the Claude 5.5 family. It accepts text, image and PDF input and returns text, with a context window of 1,000,000 tokens and an output limit of 128,000 tokens. It is the first Haiku-class model with an adjustable effort setting tied to adaptive thinking, offering low, medium and high effort levels, and thinking is on by default.

The model is built for fast, everyday work: summarization, classification requests, database queries, tagging messages, conversation compaction, live customer support, in-app assistants, lead triage, voice agents and browser use. Computer use and browser use are supported in beta through the Claude Python and TypeScript SDKs. It is also designed to run as a subagent under Claude Opus 5.5 and Claude Sonnet 5.5.

Anthropic describes it as the cheapest, fastest and most capable small model to date. It outperforms GPT-6 Luna on every benchmark where both models appear and shows strong gains over Claude Haiku 4.5 in coding, computer use and knowledge work. In the HubSpot evaluation it completed tasks faster than every model tested, with the highest hit rate and lowest false-positive rate. It is better suited to fast everyday work than to long autonomous coding sessions, where Claude Sonnet 5.5 remains clearly superior. Note that its new tokenizer counts the same text as about 30% more tokens than Claude Haiku 4.5.; 1M context
- [SpaceXAI: Grok Imagine Video 1.5 Lite API](https://openapi.felo.ai/models/x-ai/grok-imagine-video-1.5-lite): grok-imagine-video-1.5-lite; xAI; Grok Imagine Video 1.5 Lite is a distilled video generation model from SpaceXAI (xAI), positioned as the cost-performance entry level in the Grok Imagine Video 1.5 family. It accepts text and image inputs and produces video with native, synchronized audio—including sound effects, ambient sound, and lip-synced dialogue. Clips can run from 1 second to 15 seconds at 480p, 720p, or 1080p; 1080p output is rendered at 720p and then upscaled. Seven aspect ratios are supported: 16:9, 9:16, 1:1, 4:3, 3:4, 3:2, and 2:3.

The model is built for fast, affordable iteration. It is faster and lower-cost than its parent, Grok Imagine Video 1.5, making it practical to discard failed takes and regenerate. It suits drafting and prototyping through repeated trial and error, background footage and B-roll, product concept and scene ideation, and high-volume short video production where reshoots are acceptable. The safest workflow is to treat the first generation as a draft, confirm the prompt and source image work, then do the final render.

Trade-offs are notable: image quality is lower than the full model, 1080p is upscaled from 720p so detail may differ from native 1080p, subject identity may shift in long clips, and excessive motion instructions can reduce quality. Audio is plausible but not fully controllable, and the model is weak in conversational lip-sync and human anatomical accuracy. It is not listed on the image-to-video leaderboard, so its animation performance lacks independent evaluation. API access is available through partners including OpenRouter, fal, and Vercel AI Gateway; returned video URLs are temporary, so files must be downloaded promptly. Safety and licensing requirements should be checked for commercial use.
- [Nano Banana 2.1 API](https://openapi.felo.ai/models/google/gemini-nano-banana-2.1): gemini-nano-banana-2.1; Google; Nano Banana 2.1 is Google's image generation and editing model in the Gemini 3 series, built on Gemini 3.6 Flash. It accepts text, image, video, and PDF input and produces both image and text output, with a context window of 131,072 tokens and an output limit of 32,768 tokens. It supports 1K, 2K, and 4K image resolutions (default 1K) and a wide range of aspect ratios from 1:1 through 21:9.

The model emphasizes precision and control for professional-grade image creation. Mask-based editing changes only the region specified by a mask while preserving the surrounding composition, and subject consistency holds across different scenes, lighting, and angles for faces and identity. Each prompt can use up to 14 reference images, and identity consistency is maintained for up to 4 characters and 10 objects. Grounding with Google web and image search is available, thinking levels can be set to minimal, medium, or high (default medium), and conversational multi-turn editing supports iterative workflows.

It is suited to rapid iterative editing, posters and complex diagrams with clear text, historically accurate scenes and infographics that draw on long-context knowledge, localized text rendering across languages, multi-image fusion for product and multi-character scenes, and virtual try-on or character generation. Known limitations include poor small text at 1K resolution, imperfect character consistency between input and output, partial instruction-following in mask or doodle editing, occasional pose preservation and left-right spatial mix-ups, and limited world knowledge and 3D reasoning. It does not support temperature, topP, topK, seed, or logprobs parameters, nor audio generation, caching, code execution, file search, function calling, or structured outputs.; 65.536K context
- [Seedream 5.0 Flash API](https://openapi.felo.ai/models/bytedance-seed/seedream-5-0-flash): seedream-5-0-flash; ByteDance Seed; ByteDance Seed: Seedream 5.0 Flash is an image generation and editing model from the Seedream 5.0 series. It accepts text and image inputs, including up to 10 multi-reference images for composition, and produces image outputs. The model is built for speed and cost efficiency, making it suitable for latency- and cost-sensitive workloads. It supports precise interactive editing: users can circle, scribble, box, or point-and-drag to mark regions, and only those marked areas are modified while edits merge naturally back into the original scene.

The model renders native multilingual text directly in images and can decompose finished images into up to 16 separately editable layers, including subject, background, text, and decorative elements. Each layer is returned as a PNG with transparency, along with position, stacking order, and a content description. Output resolution ranges from 1K to 2K; inputs above 2K return 2K, or auto mode matches the input resolution.

Seedream 5.0 Flash is suited to tasks that demand speed and volume, such as advertising banner creation with multiple versions, e-commerce product catalogs with high-volume product images and variants, virtual try-on with precise edits to clothing or accessories, editorial and print publishing with rapid design iteration, app and UI mockup design, rapid iteration on game art, high-volume variant generation, and fast drafts or visual exploration before finalizing.
- [FLUX 3 Image API](https://openapi.felo.ai/models/black-forest-labs/flux-3-image): flux-3-image; Black Forest Labs; Black Forest Labs: FLUX 3 Image is part of the FLUX 3 multimodal foundation model family announced on July 23, 2026. It generates images from text prompts and edits images from an image plus a prompt. The model supports multi-reference generation with up to 10 reference images and produces native 4K resolution output. It offers pixel-level precise editing and uses bounding boxes to control element placement and limit edit regions. Rather than short prompts, it accepts full scene captions, and it pairs with a 9B prompt upsampler model that expands requests into detailed scene descriptions via system prompt.

The image transformer is provided as a 5.6B-parameter open-weight model, with a model size of 6B parameters in BF16 tensor type. It uses Qwen3-VL-4B-Instruct as the text encoder. An undistilled base version (Klein Base) is available for fine-tuning, LoRA training, and research, while a distilled Klein version is offered for generation, editing, and self-hosting. The checkpoint behind the BFL API is the distilled version. It integrates with the Diffusers library from day one and runs on RTX 4070-class GPUs.

FLUX 3 Image is suited to text-to-image generation, single-reference image editing, multi-reference image composition, design and advertising production, and product visualization. According to internal evaluations, FLUX 3 Image Klein placed first in text-to-image, multi-reference generation, and image editing rankings. It improves over FLUX.2 Klein 9B by 91 Elo points in text-to-image and 70 Elo points in image editing, while reducing parameter count by 38%. It outperforms larger open-weight models such as Qwen-Image 2.1 (7B parameters) and Krea 2 (12.9B parameters), and surpasses closed models Ideogram V4.5 and FLUX.2 Pro. Note that the flux-kommunity-license imposes conditions on commercial use.
- [HeyGen Video API](https://openapi.felo.ai/models/heygen/heygen-video-1): heygen-video-1; HeyGen; HeyGen Video is an AI video generation model released by HeyGen in September 2026. It generates scene visuals, dialogue and sound effects in a single pass, turning a prompt written like a film shot list — covering subject and action, camera, look, sound and constraints — into a finished clip. The model accepts text, image and reference video input and delivers video, dialogue audio and sound effect output. It offers four visual styles (claymation, live action, anime and screen print), supports 16:9, 9:16 and 1:1 formats, and renders at resolutions from 480p to 2K. Seed values and webhooks are supported for controlled, automated workflows.

HeyGen Video draws on HeyGen's avatar technology, with 700+ stock avatars and custom-built digital twins, and supports 175+ languages and dialects. In September 2026, professional voice cloning was added to the API, training a voice from 20+ minutes of a single speaker's recording. HyperFrames lets AI agents create video using HTML with deterministic rendering, frame capture, video playback and preview parity. As of 2026, the model ranked #1 on the Artificial Analysis video leaderboard, and it can be integrated flexibly via the OpenRouter API by changing the model slug.

The model is suited to personalized sales outreach, consistent social media production, brand video ads through AI product placement, multilingual localization of course materials within 24 hours, efficient training material production, and interactive tutors, coaches and support agents. Customer results illustrate the speed gains: Happy Cats localized 56 course videos into 7 languages without reshooting, School of AI grew subscribers from 10,000 to 2,000,000 in 8 months, and Tim Stacey cut video production from 2 hours to 10 minutes.
- [Ideogram 4.5 Generate API](https://openapi.felo.ai/models/felo/ideogram-4-5-generate): ideogram-4-5/generate; Ideogram; Ideogram 4.5 is an image generation and editing model from Ideogram. It accepts text and image inputs and produces image outputs. Text-to-image generation works at 1K and 2K presets, with native 2K output and supported dimensions reaching roughly 3K. The model supports prompts of up to 10,000 characters and offers four quality modes. Aspect ratios include 1:1, 3:2, 16:9, 9:16, 2:3, 3:4, and 4:5.

Editing is a core strength. It edits a source image using up to four additional reference images, optionally with a mask so changes land only where intended—mask-based inpainting preserves white pixels while editing black pixels. The Precise Edit variant applies a natural-language instruction to a source image and restores unchanged pixels from the source, keeping the source dimensions and reducing unwanted visual drift. It handles high-resolution images without downsizing first—demonstrated on a 4,016 by 6,016 source of 24.2 megapixels—and can edit a crop and merge the result back into the original. It also alters stylized lettering in place, translates copy while retaining layout, and supports pose control with a pose skeleton reference and joint angles in the prompt.

This makes Ideogram 4.5 suitable for tasks where wording and fidelity matter: in-image copy changes, product campaign variants that preserve a product's appearance across turns, architectural visualization with swapped furnishings and finishes, localization of headlines and labels, and targeted high-resolution edits. It is proprietary, with an open-weight release planned but not yet available.; 2.048K context
- [Ideogram 4.5 Precise Edit API](https://openapi.felo.ai/models/felo/ideogram-4-5-precise-edit): ideogram-4-5/precise-edit; Ideogram; Ideogram 4.5 Precise Edit is an image-editing model from Ideogram, released on September 30, 2026. It accepts text and image input and produces image output, focusing on precise local edits that alter only the requested area. Optional masks and reference images help constrain each change, and untouched regions remain pixel-identical across up to eight sequential edits without cumulative drift.

The model handles text editing, adding and removing features, and full scene changes. Prompts can be up to 10,000 characters, quality settings range from very low to high, and each request can return 1 to 8 images with a seed range. Copyright detection options are also available. Because it follows text-editing prompts exactly without adding unrequested copy, it suits product photography, interior design, photo restoration, in-image text editing, ad revisions, product colorway changes, localized copy replacement, element-by-element swaps, and label or compliance-copy correction.

For spatially complex edits involving several coordinated objects or many references, FLUX 3 may be the stronger first workflow. For a single large edit such as rebuilding a scene or changing style, whole-image redraw models may be more suitable. Ideogram 4.5 Precise Edit documents a crop-and-stitch pattern for high-resolution print masters.; 2.048K context
- [Ideogram 4.5 Replace Background API](https://openapi.felo.ai/models/felo/ideogram-4-5-replace-background): ideogram-4-5/replace-background; Ideogram; Ideogram 4.5 Replace Background is a precise editing tool in the Ideogram 4.5 family, built to remove and swap image backgrounds while preserving key foreground subjects. It delivers production-ready outputs with clean structure and sharp detail, and it preserves untouched pixels exactly during edits without pixel shifts or color changes. Untouched pixels are copied from the source image, and the original size is kept. Generation is native 2K resolution. A single request accepts up to five source images for editing, and an optional mask can limit edits to specific regions, where black pixels are edited and white pixels are preserved.

The tool suits work where the subject must stay exact while the setting changes. That includes clean product cutouts for online store listings, background changes for marketing campaigns, swapped room settings for interior design, and photo restoration workflows with sequential edits. Editing stays artifact-free across multiple rounds, and the model carries strong text rendering inherited from the Ideogram family, fitting into a comprehensive editing workflow.

It is available through the Ideogram platform and API, and through launch partners Picsart, Runway, Pika, and Leonardo AI. An open-weight release has been announced as coming soon.; 2.048K context
- [GPT-6.1 Sol API](https://openapi.felo.ai/models/openai/gpt-6.1-sol): gpt-6.1-sol; OpenAI; GPT-6.1 Sol is a GPT-6-series model from OpenAI, released at OpenAI DevDay on September 29, 2026. It sits below flagship GPT-6 Astra and above GPT-6 Luna. It accepts text, PDF, and image input and returns text output, with a 1,050,000 token context window and an output limit of 128,000 tokens. Its knowledge cutoff is April 30, 2026, and it offers data residency in the US and EU, subject to eligibility.

The model provides tiered reasoning controls from low through medium, high, xhigh, and max; reasoning cannot be fully disabled. Through the Responses API it supports function calling and tool use, structured outputs via a JSON schema in response_format, web search, file search, code interpreter, computer use, and MCP, along with hosted shell and image generation. Multi-agent functionality is in beta. It is available through the API, Codex, and ChatGPT Work for eligible users, launched in Work and Codex rather than the standard ChatGPT chat interface, and enterprise administrators may control whether it is enabled. It is served by 3 providers on OpenRouter: OpenAI, Azure, and Amazon Bedrock.

It is intended for agentic coding and multi-step software engineering, debugging, repository analysis, and code review; computer use and supervised agents that interact with applications across connected steps; document-heavy professional work with long documents and tables; multi-step business workflow automation across SaaS tools; and customer service, IT support, sales operations, research, finance operations, and business administration. It delivers near-Astra intelligence for these tasks, makes fewer factual errors than GPT-6 Sol, and respects explicit restrictions and user intent more reliably in agentic tasks.; 1.05M context
- [Claude Sonnet 5.5 API](https://openapi.felo.ai/models/anthropic/claude-sonnet-5.5): claude-sonnet-5-5; Anthropic; Claude Sonnet 5.5 is the second model in Anthropic's Claude 5.5 family, released on September 28, 2026. Built as the workhorse tier between Claude Opus 5.5 and the upcoming Claude Haiku 5.5, it accepts text and image input and returns text output, with a context window of 1,000,000 tokens and an output limit of 128,000 tokens. It supports reasoning and includes a control for thinking depth.

The model is positioned for defined, practical work: agentic tasks with clearly defined scope, bug fixes, document creation, presentations, spreadsheets, UI polishing, decks built from slide templates, and research, review and drafting. Anthropic describes it as fast and low-cost, with output speed more than 30% faster than Claude Sonnet 5 and up to 30% lower cost per task than Claude Sonnet 5. It is available through AWS, Google Cloud and Microsoft Azure, and its model identifier is claude-sonnet-5-5.

Sonnet 5.5 nearly matches Claude Opus 5.5 on benchmarks and ranks first on several of them, including Vals Index, Vibe Code Bench, Code Migration and BioMysteryBench. Its limits are worth noting: it is weaker on security tasks such as CyberBench, and Claude Opus 5.5 remains clearly better for complex open-ended work and sustained judgment. Users running Sonnet with thinking off need the new between_tools setting, and executing model-generated instructions without restriction is not recommended.; 1M context
- [GPT-6 Luna API](https://openapi.felo.ai/models/openai/gpt-6-luna): gpt-6-luna; OpenAI; GPT-6 Luna is OpenAI's efficiency-focused entry in the GPT-6 family, released September 22, 2026 as the successor to GPT-5.6 Luna. It accepts text and image input and produces text output, with a context window of 1,050,000 tokens and an output limit of 128,000 tokens. Reasoning effort can be set to none, low, medium, high, xhigh, or max, letting users balance speed against depth, and tool-enabled workflows cover functions, web search, file search, and computer use.

The model is designed for focused, high-volume work: reasoning and knowledge tasks, writing, summarization and content generation, long-document and dataset analysis in a single pass, data extraction, classification, and lightweight agentic operations. It also suits coding and developer workflows, multimodal analysis combining text and image inputs, and computer-use scenarios. Compared with GPT-5.6 Luna, it improves writing quality and computer-use capabilities.

GPT-6 Luna sits below GPT-6 Sol and GPT-6 Astra in capability and is text-only in output. It is available in the API, Codex, and ChatGPT, as well as through the Responses API, Client SDKs, Microsoft Azure, OpenRouter, and Kilo Code, and is integrated into Microsoft Foundry for testing and deployment.; 1.05M context
- [Claude Opus 5.5 API](https://openapi.felo.ai/models/anthropic/claude-opus-5.5): claude-opus-5-5; Anthropic; Claude Opus 5.5 is Anthropic's flagship model in the Claude 5.5 generation. On the Artificial Analysis Intelligence Index it ranks #1 of 352 models, making it the highest-scoring model on that leaderboard. The model supports adjustable reasoning effort, and its index score scales with that setting: 42.3 at low effort, 51.2 at medium, 53.6 at high, 56.0 at xhigh, and 57.6 at max.

Because it sits at the top of the Intelligence Index, Claude Opus 5.5 is best suited to demanding reasoning work where answer quality matters more than speed or cost — for example complex analysis, multi-step problem solving, and tasks that benefit from extended deliberation. Lower reasoning effort settings offer a way to trade some of that depth for faster responses on simpler requests.

On the Artificial Analysis Intelligence Index leaderboard, Claude Opus 5.5 is listed above Claude Sonnet 5.5, Kimi K3.1 Pro, Claude Fable 5.1, Qwen3.8 Max, GPT-6 Astra, Gemini 4 Argon (High), GPT-6.1 Sol, and Claude Opus 5.; 1M context
- [MiMo-V2.6-Flash API](https://openapi.felo.ai/models/xiaomi/mimo-v2.6-flash): mimo-v2.6-flash; Xiaomi; MiMo-V2.6-Flash is a multimodal model from Xiaomi, released on September 21, 2026 under the MIT open-source license. It uses a mixture-of-experts architecture with 309 billion total parameters and about 15 billion activated per inference. The model accepts text, image, video, and audio input and produces text output, with a context window of 1M tokens and an output limit of 131K tokens. It supports tool calling, JSON structured output, streaming, thinking mode, and compatibility with OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages APIs. Model weights are available on Hugging Face, and a technical report and reinforcement learning research resources have been released.

In the MiMo-V2.6 series, Flash is positioned as the best balance of intelligence, efficiency, and cost. Its output speed is up to 20 times faster than the previous generation, and it outperforms MiMo-V2.5-Pro on coding agent benchmarks. It is close to Pro-version performance at lower cost. The model is particularly strong on OCR text recognition, object counting, and visual reasoning. One limitation is that it may occasionally miss small visual details, such as a tilted bottle cap. In internal testing, the response-level repetition rate exceeded 0.05% and has since been optimized; 12.7% of the reinforcement learning budget was spent on scoring.

MiMo-V2.6-Flash is designed for high-frequency professional office calls, large-scale tasks, coding work, agentic tasks, and visual understanding. Its combination of broad modality support, long context, and API compatibility makes it suitable for production deployments that need efficient, cost-effective multimodal reasoning.; 1.05M context
- [SpaceXAI: Grok 4.7 API](https://openapi.felo.ai/models/x-ai/grok-4.7): grok-4.7; xAI; Grok 4.7 is SpaceXAI's flagship frontier model, succeeding Grok 4.6. It is built on a new and larger base model and was trained with a longer reinforcement learning run on a harder mix of tasks. The model accepts text input and produces text output, with a context window of 500,000 tokens and an output limit of 450,000 tokens. It supports function calling, tool choice control, structured outputs with JSON schema validation, and extended reasoning tokens, and its reasoning effort is configurable from low to xhigh. It also natively understands the Grok Bot harness.

The model is positioned as SpaceXAI's most capable to date for software development, long-running professional tasks, and general knowledge work. It is designed for long-running software engineering tasks and agentic knowledge work, including multi-hour office work such as documents and presentations, as well as professional tasks in law, finance, and electrical engineering. According to the company, it verifies its own work more carefully, handles long context more reliably, and remains effective across multi-step assignments.

Grok 4.7 ships with an entirely new safeguard stack that the company describes as its strongest model tested on refusals and jailbreak resistance, and it leads on utility for benign tasks and safe refusal on dangerous ones in dual-use domains such as cybersecurity and biological work. Its performance is sharply uneven across domains, with weaker results in clinical work. Published benchmark results were produced at the highest reasoning effort setting.; 500K context
- [DeepSeek V4.1 Flash API](https://openapi.felo.ai/models/deepseek/deepseek-v4.1-flash): deepseek-v4.1-flash; DeepSeek; DeepSeek V4.1 Flash is a multimodal mixture-of-experts model released by DeepSeek-AI on September 10, 2026. It accepts text and image input and produces text output, with native visual understanding built into its design. Its backbone holds 552 billion parameters, and an asymmetric Causal Encoder-Decoder architecture activates 8 billion parameters for input processing and 16 billion for output generation. The model is released as open weights under the MIT license and offers a context window of one million tokens with an output limit of 384,000 tokens.

The model is aimed at input-heavy agent systems where cost is dominated by large inputs such as codebases, logs, and documents, and at tasks that require processing both text and images at scale. It also suits flexible workloads scheduled during off-peak hours. Its KV cache needs only one quarter of the HBM memory and one eighth of the SSD storage of the previous generation, and deployment options include large-scale setups with 2,000 GPUs and storage clusters.

DeepSeek reports benchmark results ahead of flagship models including DeepSeek V4 Pro. It ranked first in OpenCode usage with 30 percent of observed two-month volume, reached 3 million unique users with a weekly retention rate of 69 percent, and grew daily token volume from 3 trillion on September 10 to a peak of 8.9 trillion on September 21, 2026. Its knowledge cutoff is May 2025, and it is available through the DeepSeek API and on Hugging Face. A later revision, DeepSeek V4.1 Flash 1017, is a minor update improving reasoning, coding, agentic tasks, and visual understanding.; 1.049M context
- [GPT Image 2.5 Sunburst API](https://openapi.felo.ai/models/openai/gpt-image-2.5-sunburst): gpt-image-2.5-sunburst; OpenAI; GPT Image 2.5 Sunburst is OpenAI's premium, precision-oriented tier of the GPT Image 2.5 family. It generates images from text descriptions and edits existing images with instruction-following control, accepting one image for editing or multiple images to combine styles, subjects or references. Reference images are processed at high fidelity automatically, and the model maintains consistency across multiple editing passes.

Sunburst is tuned for targeted changes without reinterpreting the entire image. It preserves identity, composition and lighting while adjusting specific elements, offers strong style control across different subjects, follows editing instructions more precisely and maintains quality over long editing sessions. It supports six quality settings (low, medium, high, xhigh, max and auto, with auto as default), multiple aspect ratios including square 1:1, portrait 9:16 and landscape 16:9, images up to 3840 pixels, plus transparency, reframing and seed control.

The model suits precision work such as object removal, lighting adjustments, scene compositing, style transfer, text translation in images, virtual clothing try-ons and multi-step image refinement systems. It also handles infographics, logos, UI mockups, photorealistic scenes, comic strips, marketing visuals and product mockups. Known limitations include noise artifacts in certain outputs; it does not support streaming, function calling, audio or video, and it costs more with longer generation times than GPT Image 2.5 Flare.
- [GPT Image 2.5 Flare API](https://openapi.felo.ai/models/openai/gpt-image-2.5-flare): gpt-image-2.5-flare; OpenAI; GPT Image 2.5 Flare is OpenAI's speed-oriented image generation and editing model in the GPT Image 2.5 series. Released on September 9, 2026, it accepts text and image inputs and produces image outputs only. It supports up to 16 reference images per request for editing, more than 9 aspect ratios including 1:1, 3:2, 2:3, 4:3, 3:4, and 16:9, and can generate up to 10 images per request. Six quality settings are available: low, medium, high, xhigh, max, and auto. Output resolutions include 1K, 2K, and 4K, and transparent backgrounds are supported.

Flare is built for speed-sensitive and high-volume work. It maintains image quality comparable to GPT Image 2 while generating more than 50% faster, and OpenAI positions it as the default choice for most applications. In third-party measurement it achieved approximately a 7x speedup over GPT Image 2, with a median of 16.7 seconds to generate a 1024x1024 image compared with 117.8 seconds for GPT Image 2. OpenAI states a 50% latency reduction. These characteristics make it well suited to interactive use and high-frequency generation.

Typical tasks include high-volume everyday image generation, creator content, rapid visual prototyping, social media image creation, frequent campaign variation creation, and everyday image generation for apps and websites. It is also useful for exploring a broad range of directions before final refinement with Sunburst. Reports note weaknesses related to image noise artifacts, and precision-oriented final editing is better handled by Sunburst. The model is available via the dedicated Images API or as an image generation tool in the Responses API, and it accepts the same parameters as Sunburst, so switching between them requires only changing the model name.
- [GPT-6 Astra API](https://openapi.felo.ai/models/openai/gpt-6-astra): gpt-6-astra; OpenAI; GPT-6 Astra is OpenAI's next-generation model, released on September 3, 2026. It is built on a computer-operator design rather than a traditional chatbot, allowing it to work across websites, desktop apps, and internal tools without APIs. It accepts text and image input and produces text output, with a context window of 1,050,000 tokens and an output limit of 128,000 tokens. OpenAI positions it as the most intelligent and most aligned model in the world.

The model is designed for advanced reasoning with deep multi-step problem solving and decision support, as well as agentic workflows that plan, execute, and coordinate multi-step tasks using tools and long-running workflows. Its long-context understanding supports large documents, repositories, and knowledge bases. It also offers advanced coding and engineering capabilities, knowledge work and content creation, extended browsing and connected intelligence, computer use and automation, and enhanced safety and alignment with updated safety training, bio/cyber safeguards, and improved monitoring protections.

GPT-6 Astra is best suited to complex reasoning and planning, software development and repository analysis, research and knowledge discovery, enterprise productivity and knowledge management, document and content creation, workflow automation with tools and connected systems, computer use across applications, and data analysis and reporting. It is distributed first to a limited set of organizations, then to ChatGPT Plus, Pro, Business, and Enterprise users, and is available through the OpenAI API, Microsoft Azure, and AWS Bedrock. Limitations include no video input and high cost; Claude Fable 5.1 outperforms it on Humanity's Last Exam with tool use and on the Artificial Analysis Intelligence Index.; 1.05M context
- [Gemini 3.8 Flash API](https://openapi.felo.ai/models/google/gemini-3.8-flash): gemini-3.8-flash; Google; Gemini 3.8 Flash is Google's most intelligent Flash model, positioned below the flagship Gemini 4 Argon. It succeeds Gemini 2.5 Flash and Gemini 3.7 Flash and is built on Gemini 3.7 Flash. It accepts text, image, video, audio, and PDF input and produces text output, with a context window of 1,000,000 tokens and an output limit of 65,536 tokens. It is a reasoning model that uses extended thinking and chain-of-thought, with a thinking level customizable across low, medium, and high.

The model supports caching, code execution, computer use (preview), file search, function calling, Google Maps grounding, search grounding, structured outputs, URL context, the Batch API, and flex and priority inference. It is aimed at long-horizon software engineering, autonomous agents, complex enterprise workflows, multi-step orchestration, and full-stack code refactoring. It also suits general reasoning, document-heavy long-running workflows, financial analysis, legal workflows, biology research, and video understanding. Its knowledge cutoff is March 2026, with some domains limited to January 2025 knowledge.

On benchmarks it scores 41 on the Artificial Analysis Intelligence Index, where the median for the same price tier is 27. It reaches 73.7% on DeepSWE v1.1, 61.4% on Vals Finance Agent v2, and 89.4% on Terminal-bench 2.1. Limitations include potential hallucination common to base models, occasional latency and timeout issues, heavy token use at high thinking levels, and a time to first token of 24.90 seconds versus a same-tier median of 3.96 seconds. It does not support audio generation, image generation, or the Live API.; 1.049M context
- [H3 Max API](https://openapi.felo.ai/models/minimax/hailuo-3-max): hailuo-3-max; MiniMax; MiniMax: H3 Max appears in the model catalog as an entry identified by its display name, MiniMax: H3 Max. The available materials contain no additional specifications, so its inputs, outputs, context window, and intended tasks cannot be described here.

No benchmarks, use cases, or capability notes accompany this entry. As a result, this description cannot recommend specific workloads or comparisons with other models in the same family.
- [Claude Fable 5.1 API](https://openapi.felo.ai/models/anthropic/claude-fable-5.1): claude-fable-5-1; Anthropic; Claude Fable 5.1 is Anthropic's flagship model, released on September 1, 2026. It accepts text and image input and returns text output, with a 1-million-token context window and support for up to 128K tokens of output. Thinking is always adaptive, with the default effort level set to high. It is generally available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS.

The model is built for demanding reasoning, long-horizon agentic work spanning multiple days, scientific discovery, and enterprise workflows involving documents, spreadsheets, and slides. It also handles implementing features across an entire codebase, code review, and performance improvement. Anthropic reports state-of-the-art performance in coding and knowledge work, with Terminal-Bench 4.0 rising from 42.0% on Claude Fable 5 to 55.8%. It also scores 35.03% on RSI-Index, ranking #1, and its writing style is described as more natural and less stereotyped than earlier Claude models.

Some trade-offs are worth noting. Output speed is 70.1 tokens per second, which is on the slower side, and the model is classified as relatively high-latency. The reported refusal rate is substantially higher than earlier Claude models, though cybersecurity-related refusals are reportedly down about 60%. Migrating from Fable 5 involves 4 breaking changes, including forced tool use now returning errors. The training data cutoff is June 2026, and retirement is not scheduled before September 1, 2027.; 1M context
- [Qwen3.8 Flash API](https://openapi.felo.ai/models/qwen/qwen3.8-flash): qwen3.8-flash; Qwen; Qwen3.8 Flash is a multimodal model from the Alibaba Qwen team that accepts text, image, video, and audio input and returns text. It uses a Mixture-of-Experts architecture with 125B main parameters, 51B N-gram embeddings, and 6B parameters activated per token. The context window reaches 262,144 tokens, with up to 262,144 tokens for reasoning content and 131,072 tokens for the final response. Thinking behavior is controllable through enable_thinking, preserve_thinking, and reasoning_effort, which offers xhigh, medium, and low levels.

The model is aimed at agentic coding and tool use, vision tasks, multi-turn agentic workflows, long video understanding, and office tasks. It also fits audio and video agent workflows such as music video creation and short drama production. Its strengths include strong reasoning and generation at notable speed, with prefill speed about 2.5 times faster thanks to the A6B architecture. It outperforms Claude Opus 4.6 Max on mobile use and real-world visual understanding tests.

Deployment options include vLLM, SGLang, TokenSpeed, and Qwen Cloud, and it can run locally via Unsloth GGUF and llama.cpp. For context beyond 262,144 tokens, RoPE scaling such as YaRN is recommended. Lower reasoning effort does not always shorten overall task completion time, since it can lead to insufficient analysis, more failures, and repeated retries.; 1M context
- [GLM 5.3 Flash API](https://openapi.felo.ai/models/z-ai/glm-5.3-flash): glm-5.3-flash; Z.ai; GLM 5.3 Flash is Z.ai's lightweight sibling to GLM 5.3, released on 26 August 2026 under the MIT license for commercial and personal use. It is a Mixture-of-Experts model with 320 billion total parameters and 18 billion active parameters per token, trained on a 30-trillion-token multimodal corpus from a newly trained base rather than reusing GLM 5.2's base. Its 45 layers sit in a hybrid architecture that combines linear and sparse attention, giving it a 4.4 times smaller KV cache and using 3.0 times less attention compute than GLM 5.3. It accepts text and image input and produces text output, with a 1 million token context window. Reasoning is on by default, and tool calling and structured output are supported. Median time to first token is 280 milliseconds and output speed is 55 tokens per second in seven-day status checks.

The model is built for agentic and multimodal work that would be expensive to run on a flagship. It targets repository-aware coding assistants, screenshot-to-code analysis, repeated code-review workflows, and document-heavy agents, including automated visual audits of PDFs, PPTXs and DOCXs. Its native GUI and interface automation support makes it usable for real-time desktop agents that click, type and read live screenshots, and it has been applied to massive refactoring and security patching of legacy codebases, converting screen recordings and designs into Next.js applications, and generating parametric CAD scripts and Blender scenes from 2D physical blueprints.

Its strengths are screenshot understanding, image analysis and visual question answering, and it beats GLM 5.2 across coding and agent tests. It should not automatically be treated as equivalent to the flagship model for the hardest reasoning or software-engineering workloads. The maximum output-token limit, cache-write price, fine-tuning support and model-specific rate limits are not prominently exposed. It is available through providers including Gate.AI, Lyceum and Z.ai's own API platform.; 1.049M context
- [Wan 3.0 API](https://openapi.felo.ai/models/alibaba/wan-3.0): wan-3.0; Alibaba; Wan 3.0 is Alibaba Tongyi Lab's video generation model, released as a public beta on August 6, 2026. Built on the diffusion-transformer paradigm refined across the Wan series, it generates native video of up to 30 seconds in a single continuous take and produces synchronized audio in the same pass. Omni-Reference lets the model accept images, video clips, audio, documents, spreadsheets, slides, PDFs, and webpage URLs, with up to 20 reference inputs per generation that can be addressed positionally in the prompt. Output supports every aspect ratio from 16:9 to 9:16 and three resolution tiers: 480p, 720p, and 1080p. It is available through Alibaba Cloud Model Studio and Qwen Cloud under the model ID wan3.0-video, and on wan.video and third-party platforms.

The model targets AI film and short-form storytelling, corporate communications and quarterly business updates, marketing explainers, short-form social content for TikTok, Instagram Reels and YouTube Shorts, e-commerce and advertising product shots, and visual restyling or scene changes on existing video. Strengths include strong visual fidelity, motion coherence, and prompt adherence, with reliable rendering of specified lighting, object placement, and action sequences. Scene consistency stays stable across the full clip length in static or slow-moving shots, and surface-detail retention holds label text and specular highlights for product shots. Document-to-video and webpage-to-video are described as an industry first.

The model is closed-weight with no local deployment option and is currently in API preview with access approval required. It requires substantial GPU resources for inference, with estimates suggesting 4 to 8 A100 or H100 GPUs for the full checkpoint. Dynamic scenes with multiple interacting subjects can show drift in secondary character positioning, and independent benchmarks are still limited. The public beta does not support 4K output.
- [Seedream 5.0 Lite API](https://openapi.felo.ai/models/bytedance-seed/seedream-5-0-lite): seedream-5-0-lite; ByteDance Seed; Seedream 5.0 Lite is the Lite variant of the Seedream 5.0 family from ByteDance Seed, released in February 2026. It balances advanced visual generation capabilities with cost efficiency, positioning itself as a smarter and more professional visual creative partner than version 4.0. The model uses a single unified multimodal architecture that combines generation, reasoning, and search. Before producing pixels, it engages in deep thinking and chain-of-thought visual reasoning over the prompt. At generation time, it can pull in live web context, current events, and trending topics through built-in real-time retrieval augmented generation with online search. An enriched world knowledge system covers multiple vertical industry knowledge bases.

The model accepts image input and produces image output. It handles typography, UI design, and chart generation as first-class use cases, and it supports visual references and complex multi-element briefs. Strengths include stronger prompt-following on complex briefs, more accurate interpretation of vague instructions, results that better conform to physical laws, and significantly improved information visualization. Compared with Seedream 4.5, it shows notable gains in knowledge reasoning, editing response, and consistency maintenance.

Seedream 5.0 Lite is suited to advertising and marketing, e-commerce operations, film and television production, digital entertainment, and education. It can generate informational posters for hot events and new product launches, hand-drawn typography and sticker designs, and app interface mockups. The model is available through Jimeng AI, the Volcano Ark experience center, Doubao for internal testing, and OpenRouter, where it has processed over 2.84 billion tokens. Note that it has been retired on some platforms such as Morphic, where users are directed to Seedream 5.0 Pro.
- [Seedream 5.0 Pro API](https://openapi.felo.ai/models/bytedance-seed/seedream-5-0-pro): seedream-5-0-pro; ByteDance Seed; No verified details are available for ByteDance Seed: Seedream 5.0 Pro. The provided facts list contains no confirmed specifications, capabilities, or benchmark results, and the supplied source links do not document this model.

Because nothing about this model has been confirmed, its suitable tasks, inputs, context length, and output limits cannot be described here. Any statement about what it can do would be speculation.

To write an accurate catalog entry, supply a verified facts list covering the model's capabilities, supported inputs, context and output limits, and benchmark results.
- [Seedance 2.0 Mini API](https://openapi.felo.ai/models/bytedance/seedance-2.0-mini): seedance-2.0-mini; ByteDance; Seedance 2.0 Mini is the lightweight, cost-oriented tier of ByteDance's Seedance 2.0 video generation family. It shares the family's unified multimodal audio-video generation architecture and prompt system, but trades output scale for speed and price: it produces4- to 12-second clips (5 seconds by default) at 480p or 720p (720p by default), in aspect ratios ranging from 16:9, 4:3, 1:1 and 3:4 to 9:16, 21:9 and 9:21.

On the input side it handles text-to-video, image-to-video, reference video, and reference audio used as background music. Up to four reference images can steer characters, avatars, clothing or environments, first- and last-frame images can anchor a shot, and avatar references can draw on trusted assets from a library. Video can be generated together with audio, while reference audio must be paired with a reference image or video.

It is built for high-throughput, cost-sensitive work: concept validation and storyboard drafts, large batches of variants, early testing of social and vertical ad creative, and scaled video asset production for e-commerce and marketing. Because output tops out at 720p and 12 seconds, a common pattern is to prototype with Mini and deliver finals with Seedance 2.0 (up to 4K and 15 seconds) or Seedance 2.5 for longer narrative pieces.
- [DeepSeek V4 Pro 0813 API](https://openapi.felo.ai/models/deepseek/deepseek-v4-pro-0813): deepseek-v4-pro-0813; DeepSeek; DeepSeek V4 Pro 0813 is DeepSeek's flagship open-weights model, released on August 13, 2026. It is a large-scale sparse mixture-of-experts model with 1.65 trillion total parameters and 49 billion activated per inference, and it has a vocabulary size of 129280. Its architecture combines hybrid compressed sparse attention with heavily compressed attention and manifold-constrained hyper-connections, while the DSpark speculative decoding module accelerates inference. It accepts text input and returns text output, supports up to 1,000,000 tokens of context with an output limit of 384K tokens, and offers both thinking and non-thinking modes, with thinking effort available at low, high, and maximum levels.

The model is built for demanding workloads: analyzing entire codebases, running multi-step automation, synthesizing large volumes of information, parsing financial research reports of 100,000+ characters, and handling cybersecurity work, software engineering agent tasks, and complex tool-calling scenarios. Its agentic capability and performance in production settings are markedly improved over earlier versions, with strong long-context stability, reasoning efficiency, and delivery reliability on complex tasks; it outperforms GLM-5.2 and approaches Opus 4.8.

DeepSeek V4 Pro 0813 was available on APP, web, and API at release. The API natively supports the OpenAI Responses API format and can be adapted for Codex with a one-click configuration script. It is also served through NVIDIA NGC with FP8 and NVFP4 quantized artifacts supporting vLLM and SGLang runtimes, and it is available on Cloudflare Workers AI and on Fireworks for on-demand deployment on dedicated GPUs with LoRA fine-tuning, with documentation from Alibaba Cloud Model Studio.; 1.049M context
- [Grok Imagine Image 2.0 API](https://openapi.felo.ai/models/x-ai/grok-imagine-image-2.0): grok-imagine-image-2.0; xAI; Grok Imagine Image 2.0 is xAI's next-generation image generation and editing model, released on August 7, 2026. It accepts text, image, and PDF input and produces images, offering a 65,536-token context window and an output limit of 58,982 tokens. The model can generate 1 to 10 images per request at 1K and 2K resolutions, with low and medium quality modes and aspect ratios including 1:1, 16:9, 9:16, 3:4, 3:2, and 2:3.

For editing, Grok Imagine Image 2.0 provides a magic wand, segmentation, background removal, and smart resize tools. It supports multi-reference editing with up to 5 reference images and can target specific regions within an image for localized edits. The model maintains consistency across repeated edits and generations, and it follows complex multi-element prompts with layout-aware generation. It is integrated into the Grok app as a quality mode and linked with X.

Grok Imagine Image 2.0 suits professional design workflows that need precise editing and layout control, marketing asset production with brand consistency across variations, and product mockups or advertising creatives using background removal and smart resize. It also supports iterative content creation with repeated small revisions, confidential and commercial creative projects, and game asset or concept art creation. Limitations include a maximum resolution of 2K, no self-hosting or fine-tuning, and some user reviews noting that GPT Image better preserves non-standard character stylization and has less censorship.
- [Seedance 2.5 API](https://openapi.felo.ai/models/bytedance/seedance-2.5): seedance-2.5; ByteDance; Seedance 2.5 is ByteDance Seed team's video generation model, announced on June 23, 2026 at the Volcano Engine FORCE conference and officially launched on July 31, 2026. It produces a single continuous shot of up to 30 seconds natively in one pass, without stitching segments. A unified joint audio-video generation architecture processes visual and audio signals in the same latent space, yielding ambient sound, foley and score that stay in sync with the generated picture. Multi-turn extension expands content to several minutes while keeping a consistent audiovisual language.

The model accepts text prompts alongside up to 50 reference inputs per generation: 30 images, 10 videos and 10 audio clips. These references can guide subject appearance, scene style, motion and voice characteristics, through white-model reference, motion reference and creative reference modes. Timestamp-level control lets creators specify actions, camera changes, rhythm and edit instructions at precise points, while region-level editing replaces a subject, background or product without re-rendering the whole clip. Green-screen, viewpoint and reference editing are also included.

Seedance 2.5 suits short dramas that need a complete narrative arc, advertising films and social media ad units, product trailers and demos, and complete onboarding videos. It also serves animated content, film and television production, corporate communications and training, and industrial, embodied-intelligence and autonomous-driving work. Its strengths are native 30-second single-shot generation that avoids character drift and lighting inconsistency from stitching, plus strong micro-level rendering quality and high controllability for brand and product fidelity.
- [H3 API](https://openapi.felo.ai/models/minimax/hailuo-3): hailuo-3; MiniMax; MiniMax H3 is an omni-modal generative model from MiniMax, released as open weights on August 3, 2026 under the MiniMax H3 Community License Agreement. It is built around the H3-Base core model, which combines a text encoder, a visual VAE, an audio VAE, and an omni transformer; RoPE is used to capture spatial and temporal relationships. The system also includes H3-Context-IR, a preprocessing and orchestration layer that is not part of the open-source release and is available only via API. A new high-compression tokenizer expands the effective sequence length fourfold and helps cut training and inference cost.

The model accepts text, image, video, and audio inputs and understands the relationships between them. From a single model it generates video and audio together, at 480p, 768p, or native 2K resolution, at 24FPS and with lengths from 5 to 15 seconds. It supports first-frame and last-frame control, image-to-video conversion, video-to-video motion transfer, voice timbre reproduction from audio references, and native stereo audio covering voice, sound effects, and music. It follows complex multimodal instructions in natural language and handles editing tasks such as lip sync to new dialogue and partial reuse of background music, while rendering on-screen text and brand expression accurately.

These abilities suit advertising, branding, e-commerce, product design, UI/UX, game content production, short-form video production with reference material, and instruction-based transitions. MiniMax describes the model as top-tier among open-source video generation models. Physical-world consistency remains a challenge, however: visual plausibility does not guarantee physical consistency, multimodal input does not always lead to effective multimodal reasoning, and reported failures include incorrect evidence grounding, incomplete event realization, and violations of contact mechanics and object geometry.
- [MoonshotAI: Kimi K3 API](https://openapi.felo.ai/models/moonshotai/kimi-k3): kimi-k3; Moonshot AI; Moonshot AI's Kimi K3 is a flagship open-weight model with 2.8 trillion total parameters, making it the first open-source model to reach the 3-trillion-parameter class. It builds on Kimi Delta Attention (KDA) hybrid linear attention and Attention Residuals (AttnRes), and uses Stable LatentMoE to activate 16 of 896 experts and about 104 billion parameters per token. Weights are released in MXFP4 format with MXFP8 activations, and the open-weight license permits use, modification, distribution, sublicensing, sale, deployment, and fine-tuning.

The model supports text, image, and video input with text output, offers a 1M-token context window, and keeps reasoning always on with low, high, and max thinking-effort settings; max is the default. It also supports tool calling and explicit prompt caching. Kimi K3 is accessible via kimi.com, the Kimi mobile app, Kimi Work desktop client, Kimi Code, Kimi API, OpenRouter, and Amazon Bedrock, where it is the first open-weight model with explicit prompt caching.

Kimi K3 is designed for long-horizon coding with minimal human supervision, understanding large codebases and terminal tools, and tasks that combine software engineering with visual reasoning such as game development, front-end engineering, and CAD. It also targets end-to-end knowledge work, production-oriented agentic workflows, and deep reasoning for frontier intelligence scenarios. Its scaling efficiency is about 2.5x better than K2's, and accuracy improved by 13 percentage points. However, it has a 51% hallucination rate, and overall performance still trails Claude Fable 5 and GPT-5.6 Sol. Deployment is demanding: the full model requires about 1.56TB of resident memory and a recommended minimum of 64 accelerators.; 1.049M context
- [GPT-5.6 Luna API](https://openapi.felo.ai/models/openai/gpt-5.6-luna): gpt-5.6-luna; OpenAI; GPT-5.6 Luna is the lightest of the three tiers in OpenAI's GPT-5.6 series, alongside GPT-5.6 Sol and GPT-5.6 Terra. In the new naming system, the number denotes the generation while Sol, Terra, and Luna denote capability tiers; Luna roughly matches the nano tier of the earlier GPT-5 family. It accepts text input and produces text output, supports tool use and multi-step workflows, and offers scalable reasoning that improves on complex tasks at higher reasoning effort. Built-in safeguards are tuned to the model's capabilities.

The model is optimized for low-latency inference, multilingual workloads, and large-context applications, with a 1,000,000-token context window. It is the fastest and most cost-efficient model in the GPT-5.6 series and performs well on high-volume, latency-sensitive tasks. Typical uses include real-time AI applications, conversational assistants, customer support, content generation, multilingual experiences, enterprise search, document understanding, summarization, workflow automation, and classification, chat, and extraction. It is not suited to complex analysis or multi-step reasoning, where it is notably weaker than Terra and Sol.

All API users can access Luna, making it the most widely available of the three tiers. ChatGPT free-tier users may access it for certain interactions, and from August 2026 it is set to become the default model for free and Go users; ChatGPT Plus users can also use it. It is available directly on Azure, deployed through Microsoft Foundry. A recommended coding split has Sol resolve uncertainty and define the plan while Luna implements specified changes, writes and runs tests, and evaluates results.; 1.05M context
- [GPT-5.6 Terra API](https://openapi.felo.ai/models/openai/gpt-5.6-terra): gpt-5.6-terra; OpenAI; GPT-5.6 Terra is a reasoning model from OpenAI, released on July 9, 2026, following a limited preview that began on June 26, 2026. It accepts text, image, and PDF/file input and returns text, with a 1.05 million token context window and an output limit of 128,000 tokens. Its knowledge cutoff is February 16, 2026. The API model ID is gpt-5.6-terra, and it is available through OpenAI, Azure, and Amazon Bedrock with automatic failover and provider routing options.

The model supports vision, function calling, prompt caching, structured output, and reasoning. Reasoning effort can be set to none, low, medium, high, xhigh, or max, with medium as the default. OpenAI describes Terra as a reasoning model built for complex, multi-step problem solving where accuracy matters more than speed. It excels at tool use and workflow execution and handles long-context understanding across large documents, knowledge bases, and multi-step workflows.

Terra is the balanced middle tier: it offers strong performance at roughly half the cost of Sol, delivers better results than GPT-5.5 while being approximately two times cheaper, and runs faster than Sol. It performs just above Fable 5 in internal evaluations and comes close to Sol when tasks have a clear, detailed system prompt and predictable structure. It suits everyday coding, code generation, agentic tasks, content workflows, business automation, knowledge retrieval, document analysis, and operational decision support.; 1.05M context
- [GPT-5.6 Sol API](https://openapi.felo.ai/models/openai/gpt-5.6-sol): gpt-5.6-sol; OpenAI; GPT-6.1 Sol is an upgrade to GPT-6 Sol and sits below the flagship GPT-6 Astra in OpenAI's GPT-6 series. It accepts text, PDF, and image input and produces text output, with a context window of 1.1M and an output limit of 128,000 tokens. It supports function calling through tools and tool_choice, structured outputs via a JSON schema in response_format, reasoning with adjustable reasoning effort settings, streaming responses, and EU in-region routing. Prompt caching applies automatically to prompts of 1024 tokens or more, and the model is served by three providers on OpenRouter: OpenAI, Azure, and Amazon Bedrock.

The model is aimed at demanding professional and agentic workloads. It targets agentic coding, computer use, document-heavy professional work, and multi-step business workflow automation. According to the provided material, it approaches Astra-level results on its target tasks at a much lower cost, and compared with GPT-6 Sol it makes fewer factual errors. It is also described as more reliable at respecting explicit restrictions and user intent during agentic tasks.

In short, GPT-6.1 Sol is the practical middle option in the GPT-6 lineup: it offers strong agentic and document-handling behavior without the flagship positioning of GPT-6 Astra, making it a fit for teams that need dependable multi-step automation and coding assistance at scale.; 1.05M context
- [Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image) API](https://openapi.felo.ai/models/google/gemini-3.1-flash-lite-image): gemini-3.1-flash-lite-image; Google; Nano Banana 2 Lite is Google's lightweight image generation and editing model, powered by the Gemini 3.1 Flash-Lite Image architecture. It accepts text, image, and video input and produces text and image output, including interleaved text and image generation. Images are generated at 1K resolution (1024x1024 pixels) across a discrete set of 14 aspect ratios, with legible text rendered directly into the image. It supports text-to-image generation as well as image editing through multi-turn conversations, handling interleaved generation and editing in a single workflow, and can accept up to 10 reference images for editing tasks.

This model is designed for speed and efficiency. It can generate a text-to-image output in as little as four seconds and maintains high character alignment matching original Nano Banana standards. It is the recommended replacement for gemini-2.5-flash-image. Every generated image carries an invisible SynthID digital watermark and C2PA metadata, both always on and not configurable. It is drop-in compatible with OpenAI APIs and supports configurable model selection within the same node.

Nano Banana 2 Lite suits high-volume 1K image generation, rapid concept previews, simple product-image variations, background or layout ideation, and latency-sensitive or cost-sensitive workloads. It is also useful for visual drafts, marketing concepts, product scenes, social media assets, thumbnails, and localized ad variations with in-image text rendering. Its knowledge cutoff date is January 2025, and it does not support 2K or 4K resolutions or Google Search grounding.; 65.536K context
- [Nano Banana Pro (Gemini 3 Pro Image) API](https://openapi.felo.ai/models/google/gemini-3-pro-image): gemini-3-pro-image; Google; Nano Banana Pro is the official name of Gemini 3 Pro Image, Google's flagship model for image generation and editing. Built on the Gemini 3 Pro base model with reasoning capability, it accepts text and image inputs and produces image outputs. It generates images at resolutions up to 4K (5632 x 3072 pixels) across four output levels (0.5K, 1K, 2K and 4K) and supports 10 image aspect ratios. The model is accessible via the Gemini app, Google AI Studio, Google Ads, Workspace and the Gemini API.

A defining strength is text rendering: Nano Banana Pro renders legible text in multiple languages, from short slogans to long paragraphs, and understands typography, fonts and calligraphy design elements. It also supports multilingual text generation and localization translation, making it suited to text-dense visual content such as posters, ads and infographics. For composition work, it can combine up to 14 reference images in a single output and maintain appearance consistency for up to 5 people.

Creative controls cover camera angle, lighting, style, color tone and focus. The model integrates with Google Search grounding to visualize real-time information such as weather and sports results, and it supports multi-turn conversational image editing. Every generated image carries an invisible SynthID digital watermark. Typical applications include marketing and advertising assets at scale with brand guidelines, infographics and charts, presentation visuals, product prototypes and concept design, educational materials, product photography and complex scene composition. Compared with Nano Banana, it offers higher image clarity, lighting realism and composition understanding, with sharper and more reliable results on product photography, complex scene composition, character consistency and text-dense visuals. Generation is slower than Nano Banana 2: about 30.8 seconds per image versus about 15.5 seconds.; 131.072K context
- [Nemotron 3 Ultra (free) API](https://openapi.felo.ai/models/nvidia/nemotron-3-ultra-550b-a55b:free): nemotron-3-ultra-550b-a55b:free; NVIDIA; Nemotron 3 Ultra is NVIDIA's final and best model in the Nemotron 3 family: an open frontier-reasoning and orchestration model built on a Hybrid Transformer-Mamba mixture-of-experts architecture. It has 550 billion total parameters with 55 billion active parameters. It uses LatentMoE for improved accuracy and includes MTP layers for faster inference through native speculative decoding. The model was pretrained in NVFP4 format and post-trained with Supervised Fine Tuning, Reinforcement Learning, and Multi-teacher On-Policy Distillation. It takes text input and returns text output, with a 1 million token context window.

The model is designed for multi-step reasoning and planning, agent orchestration, coding agents, deep research, and complex enterprise tasks. It supports function calling through tools and tool_choice parameters, structured outputs via a JSON schema in response_format, and inference time reasoning budget control. NVIDIA positions it for long-running autonomous agent workflows and high-throughput inference in high-volume agent pipelines. It outperforms state-of-the-art open LLMs on RULER at 1M context length. The free endpoint is rate limited.

NVIDIA releases open weights, training data, and recipes, including pre-trained, post-trained, and quantized checkpoints along with training datasets. On OpenRouter the model is served by three providers: DeepInfra, Baseten, and Venice.; 1M context
- [Kling: Video v3.0 Pro API](https://openapi.felo.ai/models/kwaivgi/kling-v3.0-pro): kling-v3.0-pro; Kwaivgi; Kling Video v3.0 Pro is the highest-quality tier in the Kling V3.0 family, released in February 2026. It generates video from text prompts or reference images and produces native audio, including speech, sound effects, and ambient sound, together with the video. The model supports multi-shot generation with up to 6 shots per request, each with its own prompt and duration, while automatically adjusting camera angles and composition to maintain consistency across scenes. It also supports start-frame and end-frame specification for controlled transitions, negative prompts to exclude unwanted elements, and an element list feature to keep specified visual elements consistent.

The model maintains character identity in scenes with 3 or more characters through multi-character coreference and supports lip-sync so character mouth movements match the audio. Its prompt enhancer automatically improves user-written motion descriptions. Audio can be generated in Chinese, English, Japanese, Korean, and Spanish, and the model reproduces Chinese dialects such as Cantonese and Sichuanese as well as accents like American English, British English, and Indian English. Clips range from 3 to 15 seconds, with landscape, portrait, and square aspect ratios.

Kling Video v3.0 Pro is designed for premium video production requiring cinematic quality and realistic detail and motion. It suits scene transition creation using start and end frames, marketing and advertising product promo videos with logo and text retention, character animation from portraits with detailed expressions and motion, and longer animations up to 15 seconds. It targets professional video creators, advertising agencies, and content creators, and is accessible via fal, WaveSpeed, Runway Dev, Kie AI, and MindStudio, as well as integrated into creative tools such as Adobe Firefly and Picsart.
- [Kling: Video v3.0 Standard API](https://openapi.felo.ai/models/kwaivgi/kling-v3.0-std): kling-v3.0-std; Kwaivgi; Kling 3.0 Standard is a multimodal generative model from Kling that produces video and audio in a single pass. Built on the Omni One architecture with 3D Spacetime Joint Attention and Chain-of-Thought reasoning, it accepts text or image inputs and outputs clips from 3 to 15 seconds long. It supports landscape, portrait, and square aspect ratios and can create multi-shot sequences of up to 6 shots per request, each with its own prompt and duration, while maintaining scene consistency across cuts. Audio—including dialogue, sound effects, and ambient sound—is generated natively alongside the video, eliminating a separate audio post-production step.

The model is designed for high-volume experimentation and rapid iteration. It is particularly suited for social media content for platforms such as Instagram, TikTok, and YouTube Shorts, e-commerce product videos, and short narrative sequences like mini commercials and story-driven posts. Physics-aware motion modeling—covering gravity, balance, deformation, collision, and inertia—helps avoid floating objects, broken limbs, and unnatural motion artifacts. Strong prompt adherence ensures high-fidelity interpretation of text and visual prompts, and semantic consistency is maintained across generations.

As the entry-level tier of the Kling 3.0 family, Standard renders video at 720p resolution. This is lower than the Pro tier at 1080p and the 4K tier at the largest size. Standard is positioned as the most cost-effective option among the three Kling 3.0 tiers. Kling 3.0 is described as the first unified multimodal model generating video, audio, and images within a single architecture. The model is available through platforms including Runway and fal, with API access offered through Runway Dev using the same endpoints, fields, and durations across all tiers, differing only in output size and price.
- [GPT-5.5 API](https://openapi.felo.ai/models/openai/gpt-5.5): gpt-5.5; OpenAI; OpenAI: GPT-5.5 is OpenAI's flagship model for the most complex professional work, released on April 23, 2026. It accepts text, images, audio, and video within a single unified architecture, and produces text output. With a 1 million token context, it can handle very long documents and complex multi-turn conversations.

The model is built for agentic work: it autonomously plans and completes multi-step tasks across multiple tools, uses tools more effectively, and continues work until completion. It understands task intent faster and waits for user instructions less often. It also supports multilingual use and vision. GPT-5.5 is more token-efficient than GPT-5.4, achieving better results with fewer tokens.

GPT-5.5 suits complex customer case handling such as refund disputes and multi-policy questions, document-heavy analysis including contract review, report generation, and structured data extraction, research and information synthesis across multiple sources, long multi-day knowledge work and scientific research workflows, and operating real software interfaces. A Pro variant uses parallel test-time compute. It is classified as High capability in cybersecurity but not Critical, and was not evaluated as able to develop functional zero-day exploits without human intervention.; 1.05M context
- [Veo 3.1 Fast API](https://openapi.felo.ai/models/google/veo-3.1-fast): veo-3.1-fast; Google; Veo 3.1 Fast is a Google first-party generative AI model built for video creation. It supports text-to-video and image-to-video generation, allowing users to produce video content from written prompts or existing images. Google positions it as a leading video generation model and a premium, high-quality tool within its AI creative suite.

The model is available in Google Cloud Model Garden, where it sits alongside other Google first-party models such as Gemini and Gemma, as well as third-party models. Model Garden provides access to more than 200 models in total. Within Google's lineup, Veo 3.1 Fast is grouped with Imagen for text-to-image and Chirp for speech-to-text as part of the company's first-party model portfolio.

Veo 3.1 Fast is suited to cinematic video creation and other tasks that require generating video from text descriptions or images. Its inclusion in Model Garden makes it accessible through the same environment Google provides for its broader collection of first-party and third-party models.
- [Seedance 2.0 API](https://openapi.felo.ai/models/bytedance/seedance-2.0): seedance-2.0; ByteDance; Seedance 2.0, released by ByteDance in February 2026, is a professional-grade video generation foundation model built on a Mixture of Experts architecture with a parameter scale of 200 billion. It produces audio and video together natively, synchronizing dialogue, sound effects, and background music with the visuals from the start. The model accepts up to 12 assets in a single generation—up to 9 images, 3 video clips, and 3 audio files—and supports 15-second high-quality multi-shot output with dual-channel audio.

The model offers intelligent duration control, where setting duration to a negative value lets it choose the best length, and it performs adaptive aspect ratio selection based on the provided inputs. It also supports video editing and extension while preserving original motion and camera work, and users can turn audio off for silent video. Seedance 2.0 is designed for film and video production, marketing and advertising, e-commerce product showcases, educational training videos, and social media short-form content.

Safety and provenance are addressed at the model level: generation of real people's likenesses, copyrighted characters, and protected brand identities is restricted, and every generated video embeds C2PA watermarking that cryptographically signs provenance metadata. Seedance 2.0 ranked first on Arena text-to-video, Arena image-to-video, Artificial Analysis image-to-video with audio, and Tsinghua University's WorldReasonBench among all tested models. It is available through CapCut and Dreamina as well as API providers including Fal and OpenRouter.
- [Seedance 2.0 Fast API](https://openapi.felo.ai/models/bytedance/seedance-2.0-fast): seedance-2.0-fast; ByteDance; Seedance 2.0 Fast is ByteDance's speed-optimized entry in the Seedance 2.0 series. It shares a unified multimodal architecture with standard Seedance 2.0 and is distilled for faster generation. It accepts text, image, video, and audio input and outputs video with native synchronized audio, covering text-to-video, image-to-video, and video-to-video generation. Clips range from 4 to 15 seconds at 480p, 720p, or 1080p.

Multi-reference input includes first-frame and last-frame control, up to 9 reference images, and up to 3 reference videos of 2 to 15 seconds each. Prompts can also edit specified segments or extend existing video. These controls make the model practical for reference-guided clips and for iterating on a direction without rebuilding each shot from scratch.

Seedance 2.0 Fast is best suited to rapid video prototyping, advertising drafts, storyboard visualization, social media creative variants, and batch creative exploration—especially workflows that generate many candidate versions before selecting a final direction. Teams should note weaker multi-subject consistency, limited text generation and complex editing, occasional audio distortion, fewer generation modes and resolution options than standard Seedance 2.0, and no 4K support.
- [Gemma 4 26B A4B  (free) API](https://openapi.felo.ai/models/google/gemma-4-26b-a4b-it:free): gemma-4-26b-a4b-it:free; Google; Gemma 4 26B A4B is an open-weight multimodal model from Google DeepMind built on a Mixture-of-Experts architecture with 26B total parameters and 4B active parameters at inference. It accepts text, image, and video input and produces text output, with a 256K token context window. Images can have variable aspect ratios and resolutions, and video is processed up to 60 seconds at 1 frame per second. The model supports over 140 languages, native system prompts, native function calling, and a configurable thinking mode.

It is designed for coding assistance in IDEs and agentic workflows, code generation, debugging and refactoring, scientific and mathematical problem solving, research support, document parsing and OCR, image-based question answering, video understanding, and large-scale document processing and RAG systems. Its strengths include a balance of intelligence against parameter efficiency, near-frontier performance, and the ability to run on consumer GPUs and workstations. Reported throughput is about 91 tokens per second with a time to first token of about 0.92 seconds, and it is said to run within 2% of the 31B dense model on every benchmark tested.

Gemma 4 26B A4B is licensed under Apache 2.0 and available for commercial use. It can be accessed through Cloudflare, Google Cloud Vertex AI, NVIDIA NGC, and Azure ML, and API access is offered by 7 providers. It does not natively support audio; audio processing is available only in the smaller E2B, E4B, and 12B models. The model is positioned as the mid-range mainstay between the edge-oriented small models and the enterprise-oriented large model in the Gemma 4 family.; 262.144K context
- [MoonshotAI: Kimi K2.5 API](https://openapi.felo.ai/models/moonshotai/kimi-k2.5): kimi-k2.5; Moonshot AI; Kimi K2.5 is an open-weight native multimodal model from Moonshot AI that understands both text and images. Built on Kimi-K2-Base through continued pretraining on about 15 trillion mixed vision and text tokens, it uses a Mixture-of-Experts architecture with 1.04 trillion total parameters and 32B active parameters per inference, plus a 400 million parameter MoonViT vision encoder. It offers a thinking mode for step-by-step analysis and an instant mode for fast responses, and supports tool calling through an OpenAI/Anthropic-compatible API.

Its agentic capabilities are a defining feature: Agent Swarm decomposes complex tasks into parallel subtasks handled by dynamically instantiated domain-specific agents, coordinating up to 100 subagents and executing up to 1,500 tool calls. Coding with Vision turns visual specifications such as UI designs and video workflows into code, and the model autonomously coordinates tools for visual data processing. Released under a modified MIT license permitting commercial use, it runs on inference engines including vLLM, SGLang and KTransformers.

The model suits multi-agent task coordination at scale, complex data processing pipelines that scrape, clean, analyze and visualize data in parallel where dependencies allow, code generation from visual specifications, and agentic workflows executing 200 to 300 tool calls autonomously. Kimi Code CLI is recommended as the agent framework for optimal performance. Among open-weight models it took first place on 13 of 17 Vals AI benchmarks, though Claude Opus 4.5 remains the leader on SWE-Bench for maximum code generation accuracy.; 262.144K context
- [Gemini 2.5 Pro API](https://openapi.felo.ai/models/google/gemini-2.5-pro): gemini-2.5-pro; Google; Gemini 2.5 Pro is Google DeepMind's flagship model, featuring a hybrid reasoning architecture that applies extended thinking before responding. It combines a 1 million token context window with multimodal understanding across text and video input, plus audio generation and real-time voice interaction. The model supports grounding with Google Search, parallel web search, code execution, and function calling, making it a versatile foundation for demanding workloads.

The model is positioned as Google's strongest offering for coding and world knowledge. It excels at web development and the creation of visually compelling web apps, agentic code applications, and code transformation and editing. Its video capabilities are state-of-the-art, including dynamic animation generation from videos with a single prompt. For complex prompts that require multiple rounds of strategizing and iterative solving, the extended thinking mode offers a structured path to higher-quality answers.

Gemini 2.5 Pro is well suited to software development, particularly web development, as well as building agentic applications and creative video work. It also handles natural real-time voice conversations and general-purpose AI tasks. Running at 110 tokens per second, it delivers strong reasoning performance on Math and STEM benchmarks and outperforms Gemini 2.5 Flash on all nine shared benchmarks.; 1.049M context
- [GPT-3.5 Turbo (older v0613) API](https://openapi.felo.ai/models/openai/gpt-3.5-turbo-0613): gpt-3.5-turbo-0613; OpenAI; GPT-3.5 Turbo (older v0613) is an older version of OpenAI's GPT-3.5 Turbo model, designed to understand and generate natural language or code. It is optimized for chat and traditional completion tasks and accepts tools and tool_choice parameters for function calling. It also supports structured outputs via a JSON schema and can be fine-tuned for specific use cases. The model has a 4,095-token context window and a 4,096-token completion limit.

The model is suited to straightforward text tasks such as classification, basic question answering, and text extraction. It can pull data from large amounts of text and power AI personal assistants for internal or customer use. Natural language understanding and generation tasks are also in scope, and fine-tuning can improve steerability or reliable output formatting. It is described as OpenAI's fastest model and is cost-effective for high-volume workloads.

Key limitations should be considered: its knowledge cutoff is September 30, 2021, and it was removed as of June 13, 2024 according to one source and is deprecated. It is not currently available through EU in-region routing. Newer versions such as gpt-3.5-turbo-0125 offer improved instruction-following, JSON mode, and parallel function calling with a default 16K context window.; 4.095K context

## Model Cases
- [JEV cases](https://openapi.felo.ai/model-cases/jev): Real-world examples of what JEV can do, with prompts and results.
- [GPT-Image-2.5 cases](https://openapi.felo.ai/model-cases/gpt-image-2-5): Real-world examples of what GPT-Image-2.5 can do, with prompts and results.
- [GPT-Image-2.6 cases](https://openapi.felo.ai/model-cases/gpt-image-2-6): Real-world examples of what GPT-Image-2.6 can do, with prompts and results.
- [Seedance 2.5 cases](https://openapi.felo.ai/model-cases/seedance-2-5): Real-world examples of what Seedance 2.5 can do, with prompts and results.
- [Seedance 2.5 4K cases](https://openapi.felo.ai/model-cases/seedance-2-5-4k): Real-world examples of what Seedance 2.5 4K can do, with prompts and results.
- [Claude Opus 5.5 cases](https://openapi.felo.ai/model-cases/opus-5-5): Real-world examples of what Claude Opus 5.5 can do, with prompts and results.
- [Claude Haiku 5.5 cases](https://openapi.felo.ai/model-cases/haiku-5-5): Real-world examples of what Claude Haiku 5.5 can do, with prompts and results.
- [Nano Banana 2.1 cases](https://openapi.felo.ai/model-cases/nano-banana-2-1): Real-world examples of what Nano Banana 2.1 can do, with prompts and results.
- [Nano Banana Pro cases](https://openapi.felo.ai/model-cases/nano-banana-pro): Real-world examples of what Nano Banana Pro can do, with prompts and results.
- [Kling 4.0 cases](https://openapi.felo.ai/model-cases/kling-4): Real-world examples of what Kling 4.0 can do, with prompts and results.
- [Grok Imagine cases](https://openapi.felo.ai/model-cases/grok-imagine): Real-world examples of what Grok Imagine can do, with prompts and results.
- [Veo 3.1 cases](https://openapi.felo.ai/model-cases/veo-3-1): Real-world examples of what Veo 3.1 can do, with prompts and results.
- [Ideogram cases](https://openapi.felo.ai/model-cases/ideogram): Real-world examples of what Ideogram can do, with prompts and results.

## Use Cases
- [Turn one question into a sourced research brief](https://openapi.felo.ai/use-cases/research-brief): A copy-paste recipe for a research agent: Chat API for a sourced answer, Web Fetch for full pages, X Search for social signal, and LiveDocs for reusable evidence.
- [Build a competitor teardown your sales team can use](https://openapi.felo.ai/use-cases/competitor-analysis): A copy-paste competitive teardown recipe: Chat API for a sourced baseline, Web Fetch for the competitor's own pages, X Search for buyer signal, and a LiveDoc you can refresh each quarter.
- [Turn a market question into a go-to-market brief](https://openapi.felo.ai/use-cases/gtm-research): A copy-paste GTM research recipe: Chat API for a sourced market read, Web Fetch for category and pricing pages, X Search for buyer language, and a LiveDoc that keeps the evidence reusable.
- [Research a product niche before you stock it](https://openapi.felo.ai/use-cases/ecommerce-product-research): A product-research workflow for ecommerce: sourced category context from Chat, competitor pages captured with Web Fetch, buyer complaints from X Search, and the evidence kept in a LiveDoc.
- [Turn one search query into an SEO content brief](https://openapi.felo.ai/use-cases/seo-content-brief): A copy-paste SEO brief recipe: Chat API for current coverage, Web Fetch for the pages that answer the query, and a LiveDoc archive for every brief. It reads public pages; it does not measure rankings or search volume.
- [Research a target account before your first outreach](https://openapi.felo.ai/use-cases/account-research): A copy-paste account research recipe: Chat API for a sourced overview, Web Fetch for the company site, pricing, and careers pages, X Search for recent posts, and a LiveDoc that keeps the evidence for the next account.
- [Ship a daily AI news digest your team actually reads](https://openapi.felo.ai/use-cases/ai-news-digest): A copy-paste recipe for an AI news digest: Chat API for a sourced roundup, X Search for practitioner posts, YouTube Subtitling for a transcript, and a LiveDoc archive of past issues.
- [Turn social chatter into a weekly listening brief](https://openapi.felo.ai/use-cases/social-listening): A copy-paste social listening workflow: X Search for dated brand and competitor posts, Chat for sourced context, Web Fetch for the pages behind it, and a LiveDoc archive of every weekly brief.
- [Turn a YouTube video into a publishable article](https://openapi.felo.ai/use-cases/video-to-article): A copy-paste recipe for repurposing video: YouTube Subtitling for the transcript, Web Fetch for linked sources, the LLM API for the outline and draft, and LiveDocs for review.
- [Fill a month of content from the questions your audience asks](https://openapi.felo.ai/use-cases/content-calendar): A content calendar workflow: Chat for themes, X Search for dated audience questions, YouTube Subtitling for a reference video, and a four-week calendar archived in a LiveDoc.
- [Turn a product brief into a landing page](https://openapi.felo.ai/use-cases/landing-page): A workflow for founders and product marketers: research the audience with Chat and Web Fetch, store the brief in a LiveDoc, generate a hosted landing page with the Landing Page Task API, and review it before sharing.
- [Make a motion-graphics video with HTML and Opus 5.5](https://openapi.felo.ai/use-cases/html-video): Use Claude Opus 5.5 through the Felo API to write a deterministic HTML scene, render every frame with Playwright, and encode a polished MP4 with ffmpeg. No Claude subscription required — just a Felo API key.
- [Turn a research page into a slide deck](https://openapi.felo.ai/use-cases/web-to-slides): A step-by-step PPT Task API recipe: extract a source page with Web Fetch, store it as LiveDoc context, submit an async slide task, and poll until the deck is ready.
- [Answer support questions from your own documentation](https://openapi.felo.ai/use-cases/support-knowledge-base): A grounded support-answer recipe: create a LiveDoc, add your help center and runbooks, retrieve the right passages, and answer with the OpenAI-compatible model endpoint.
- [Triage a support ticket into a grounded reply draft](https://openapi.felo.ai/use-cases/support-triage): A support triage workflow: redact and classify a ticket with Chat, retrieve grounded passages from a LiveDoc, check public incident chatter, and log a routed reply draft for human review.
- [Keep model spend predictable as usage grows](https://openapi.felo.ai/use-cases/model-cost-routing): A copy-paste recipe for predictable LLM spend: route each task to a fast or strong model ID on one endpoint, log the usage tokens every response returns, and read current rates from the live catalog.

## Use Cases by Team
- [Founders & operators use cases](https://openapi.felo.ai/use-cases/for/founders): 5 workflows for founders & operators: Research a market, size a decision, and get a first draft before the team exists.
- [Marketing & content use cases](https://openapi.felo.ai/use-cases/for/marketers): 10 workflows for marketing & content: Turn research, social chatter and source material into campaigns, briefs, and pages.
- [Sales & partnerships use cases](https://openapi.felo.ai/use-cases/for/sales): 4 workflows for sales & partnerships: Walk into a conversation with sourced context, a competitor angle, and a clear next step.
- [Support & success use cases](https://openapi.felo.ai/use-cases/for/support): 3 workflows for support & success: Answer from your own docs, triage the queue, and keep the knowledge base current.
- [Commerce & product use cases](https://openapi.felo.ai/use-cases/for/commerce): 2 workflows for commerce & product: Check demand, compare competing stores, and validate a product before you commit stock.
- [Builders & platform use cases](https://openapi.felo.ai/use-cases/for/builders): 3 workflows for builders & platform: Wire Felo into an agent or product, keep model spend visible, and ship a reviewable flow.

## Harness APIs
- [Felo · PPT Generation and Export](https://openapi.felo.ai/docs/data-api/guides/felo-ppt): Create a presentation, poll its progress, export a PPTX and manage its lifetime.
- [Felo · Web Search](https://openapi.felo.ai/docs/data-api/guides/felo-web-search): Find web pages and return titles, links and snippets.
- [Felo · Web Read](https://openapi.felo.ai/docs/data-api/guides/felo-web-read): Read a URL as Markdown, text, HTML or a screenshot.
- [Felo · X Data](https://openapi.felo.ai/docs/data-api/guides/felo-x-data): Retrieve X profiles, search users and posts, and paginate timelines and replies.
- [Felo · Search-grounded Chat](https://openapi.felo.ai/docs/data-api/guides/felo-chat): Generate an answer with web sources using platform balance.
- [Firecrawl](https://openapi.felo.ai/apis/firecrawl): Turn sites into clean markdown, structured data or screenshots.
- [Tavily](https://openapi.felo.ai/apis/tavily): Web search, page extraction and site crawling for agents.
- [Exa](https://openapi.felo.ai/apis/exa): Neural and keyword web search with page contents returned.
- [Ahrefs](https://openapi.felo.ai/apis/ahrefs): Use Ahrefs to retrieve domain ratings and site metrics, read competitor pages with Felo, then ask your Agent to prepare an SEO comparison and action list.
- [Image Generation](https://openapi.felo.ai/docs/guides/image-generation): Generate and edit images asynchronously: submit a task, poll for completion, and download signed output URLs.

## Aggregated APIs
- [Office APIs](https://openapi.felo.ai/apis/office): Felo's own APIs: chat with search, model calls and asynchronous document generation.
- [Marketing APIs](https://openapi.felo.ai/apis/marketing): Company and contact data for outreach and sales development.
- [Search APIs](https://openapi.felo.ai/apis/search): Web search, page extraction, site crawling and visibility in AI answers.
- [Social media APIs](https://openapi.felo.ai/apis/social): Public posts, profiles, comments, transcripts and trends from the major platforms.
- [SEO APIs](https://openapi.felo.ai/apis/seo): Backlink indexes, keyword rankings, domain analytics and AI-visibility data.
- [Finance and market data APIs](https://openapi.felo.ai/apis/finance): Crypto prices, filings, company fundamentals, market prices and prediction markets.
- [Felo API](https://openapi.felo.ai/apis/felo): Web search, search-grounded chat, web content, X data and presentation generation. Use your platform balance with a Felo API key.
- [Similarweb API](https://openapi.felo.ai/apis/similarweb): Traffic estimates, audience composition, rankings and referral sources.
- [Semrush API](https://openapi.felo.ai/apis/semrush): Domain analytics, keyword rankings, traffic estimates and competitor research.
- [Ahrefs API](https://openapi.felo.ai/apis/ahrefs): Backlink indexes, referring domains, organic keywords and site health.
- [DataForSEO API](https://openapi.felo.ai/apis/dataforseo): SEO, search-engine results, keyword research, backlinks and AI visibility data.
- [Apollo API](https://openapi.felo.ai/apis/apollo): Company and contact records for go-to-market and prospecting.
- [Creator Discovery API](https://openapi.felo.ai/apis/creator-discovery): Creator contact emails from a profile URL, and similar-creator discovery.
- [Tavily API](https://openapi.felo.ai/apis/tavily): Web search, page extraction and site crawling for agents.
- [Firecrawl API](https://openapi.felo.ai/apis/firecrawl): Turn sites into clean markdown, structured data or screenshots.
- [Oxylabs API](https://openapi.felo.ai/apis/oxylabs): Query an AI answer engine to see how a page shows up in AI-generated answers.
- [X API](https://openapi.felo.ai/apis/x): Query public users, posts, communities and trends.
- [Reddit API](https://openapi.felo.ai/apis/reddit): Subreddit posts, comment threads, user profiles and search.
- [Instagram API](https://openapi.felo.ai/apis/instagram): Public profiles, posts, reels, comments and hashtag activity.
- [Pinterest API](https://openapi.felo.ai/apis/pinterest): Public pins, boards, profiles and visual search results.
- [Exa API](https://openapi.felo.ai/apis/exa): Neural and keyword web search with page contents returned.
- [BytePlus API](https://openapi.felo.ai/apis/byteplus): Real-time web search and clean, structured page content.
- [Anthropic Web Search API](https://openapi.felo.ai/apis/anthropic-web-search): Web search with citations, using Anthropic's own search tool.
- [OpenAI Web Search API](https://openapi.felo.ai/apis/openai-web-search): Web search grounded by OpenAI's own search tool.
- [Financial Data API](https://openapi.felo.ai/apis/financial): Company fundamentals, financial statements, market prices and economic data.
- [EDINET API](https://openapi.felo.ai/apis/edinet): Japanese statutory filings: list disclosure documents or download one by ID.
- [YouTube API](https://openapi.felo.ai/apis/youtube): Video search, channel and playlist data, comments and transcripts.
- [Scholar Search API](https://openapi.felo.ai/apis/scholar-search): Academic paper search, web-and-scholar combined results, and result explanations.
- [Kalshi API](https://openapi.felo.ai/apis/kalshi): Event contract markets: prices, order books and settlement outcomes.
- [Perplexity API](https://openapi.felo.ai/apis/perplexity): Answer-engine results with sources and related questions.
- [Polymarket API](https://openapi.felo.ai/apis/polymarket): Prediction market prices, order books and resolved outcomes.
- [CoinGecko API](https://openapi.felo.ai/apis/coingecko): Cryptocurrency prices, market capitalization, exchanges and token data.

## Documentation
- [Documentation index](https://openapi.felo.ai/docs.md): One Markdown index of every documentation page, grouped like the docs sidebar.
- [Felo API Setup](https://openapi.felo.ai/docs/api-reference/openapi/overview.md): Set up Felo endpoints in Data API for search, web content, X data and presentations, using platform balance.
- [Web Search](https://openapi.felo.ai/docs/api-reference/openapi/web-search.md): Find web pages and return titles, links and snippets.
- [Web Read](https://openapi.felo.ai/docs/api-reference/openapi/web-read.md): Read a URL as Markdown, text, HTML or a screenshot.
- [X Data](https://openapi.felo.ai/docs/api-reference/openapi/x-data.md): Retrieve X profiles, search users and posts, and paginate timelines and replies.
- [Search-grounded Chat](https://openapi.felo.ai/docs/api-reference/openapi/chat.md): Generate an answer with web sources using platform balance.
- [PPT Generation and Export](https://openapi.felo.ai/docs/api-reference/openapi/ppt.md): Create a presentation, poll its progress, export a PPTX and manage its lifetime.
- [OpenAPI Billing and Errors](https://openapi.felo.ai/docs/api-reference/openapi/billing.md): Understand platform balance, request and result charges, page billing and safe retries.
- [Getting Started](https://openapi.felo.ai/docs/getting-started.md): Create one Felo key, then connect an AI tool with one command or call the API directly.
- [FeloAPI CLI](https://openapi.felo.ai/docs/cli.md): Install FeloAPI CLI with npm and run feloapi setup to configure Codex or Claude Code with your Felo API key.
- [Authentication](https://openapi.felo.ai/docs/authentication.md): Authenticate requests with API keys, follow security practices, and handle auth errors.
- [Platform Overview](https://openapi.felo.ai/docs/guides/overview.md): Choose between model APIs, third-party data, platform-balance Felo OpenAPI and Pro-credit Felo APIs.
- [Models and Pricing](https://openapi.felo.ai/docs/guides/models-and-pricing.md): Understand model IDs, providers, published rates, discounts, billing units, and how a request is priced.
- [Why does the same model have different prices?](https://openapi.felo.ai/docs/guides/provider-routing.md): Understand supplier prices, how Felo chooses a provider, and how to select one yourself.
- [Protocols and SDKs](https://openapi.felo.ai/docs/guides/protocols.md): Use OpenAI-compatible, Responses, and Anthropic Messages protocols, including streaming, tool calls, and migration.
- [Async Tasks](https://openapi.felo.ai/docs/guides/async-tasks.md): Submit long-running jobs, poll task status, handle terminal states, and retrieve results.
- [Image Generation](https://openapi.felo.ai/docs/guides/image-generation.md): Generate and edit images asynchronously: submit a task, poll for completion, and download signed output URLs.
- [Errors, Limits, and Retries](https://openapi.felo.ai/docs/guides/errors-and-limits.md): Read error responses and rate-limit headers, retry safely, and diagnose common failures.
- [Data API Overview](https://openapi.felo.ai/docs/guides/data-api.md): Call third-party endpoints through one shared prefix, with consistent authentication and request handling.
- [Felo APIs Overview](https://openapi.felo.ai/docs/api-reference/felo-apis.md): Felo workflows under /v2, called with Felo Pro credits. Compare them with platform-balance APIs.
- [Felo Skills for AI Agents](https://openapi.felo.ai/docs/felo-skills.md): Install every Felo AI Skill into Codex, Claude Code, OpenClaw, or Hermes Agent.
- [Claude Code and Codex Configuration](https://openapi.felo.ai/docs/api-reference/v2/llm-agent-clients.md): Configure Claude Code and Codex to route model calls through Felo API Platform. Model calls require prepaid API credits; Felo Pro does not apply.
- [Codex](https://openapi.felo.ai/docs/agent-clients/codex.md): Configure Codex CLI with a Felo custom model provider.
- [Claude Code](https://openapi.felo.ai/docs/agent-clients/claude-code.md): Configure Claude Code to route Anthropic-compatible model calls through Felo API Platform.
- [OpenCode](https://openapi.felo.ai/docs/agent-clients/opencode.md): Connect OpenCode to Felo through the OpenAI-compatible endpoint.
- [OpenClaw](https://openapi.felo.ai/docs/agent-clients/openclaw.md): Connect OpenClaw to Felo API Platform for assistant and agent workflows.
- [Hermes Agent](https://openapi.felo.ai/docs/agent-clients/hermes-agent.md): Configure Hermes Agent with Felo API Platform model access.
- [TRAE](https://openapi.felo.ai/docs/agent-clients/trae.md): Use Felo API Platform from TRAE through compatible model provider settings.
- [OpenViking](https://openapi.felo.ai/docs/agent-clients/openviking.md): Connect OpenViking to Felo API Platform using a compatible provider profile.
- [DeepSeek Harness](https://openapi.felo.ai/docs/agent-clients/deepseek-harness.md): Connect DeepSeek Harness to Felo through its native DeepSeek adapter or a custom OpenAI-compatible provider.
- [Cursor](https://openapi.felo.ai/docs/agent-clients/cursor.md): Use Felo API Platform in Cursor through custom OpenAI-compatible settings where available.
- [Cline](https://openapi.felo.ai/docs/agent-clients/cline.md): Configure Cline with Felo API Platform model access.
- [Pi](https://openapi.felo.ai/docs/agent-clients/pi.md): Install the Felo Skill for Pi and optionally add Felo as a Pi model provider.
- [Freebuff](https://openapi.felo.ai/docs/agent-clients/freebuff.md): Install the Felo Skill for Freebuff; it shares the universal ~/.agents/skills directory with Cline.
- [Roo Code](https://openapi.felo.ai/docs/agent-clients/roo-code.md): Configure Roo Code with a Felo-compatible model provider.
- [Kilo Code](https://openapi.felo.ai/docs/agent-clients/kilo-code.md): Configure Kilo Code with Felo API Platform model access.
- [Magpie](https://openapi.felo.ai/docs/agent-clients/magpie.md): Connect Magpie to Felo API Platform so its local gateway serves Felo models to Codex, Claude Code, and other agents.
- [Dify](https://openapi.felo.ai/docs/agent-clients/dify.md): Connect Dify to Felo Model API and Data API.
- [Coze](https://openapi.felo.ai/docs/agent-clients/coze.md): Connect Coze to Felo Model API and Data API.
- [n8n](https://openapi.felo.ai/docs/agent-clients/n8n.md): Connect n8n to Felo Model API and Data API.
- [Zapier](https://openapi.felo.ai/docs/agent-clients/zapier.md): Connect Zapier to Felo Model API and Data API.
- [Make](https://openapi.felo.ai/docs/agent-clients/make.md): Connect Make to Felo Model API and Data API.
- [Meta Muse](https://openapi.felo.ai/docs/agent-clients/meta-muse.md): Use Felo Data API search results in Meta Muse.
- [OpenAI Dots](https://openapi.felo.ai/docs/agent-clients/openai-dots.md): Use Felo Data API search results in OpenAI Dots.
- [Gemini Spark](https://openapi.felo.ai/docs/agent-clients/gemini-spark.md): Use Felo Data API search results in Gemini Spark.
- [Grok Bot](https://openapi.felo.ai/docs/agent-clients/grok-bot.md): Use Felo Data API search results in Grok Bot.
- [Manus](https://openapi.felo.ai/docs/agent-clients/manus.md): Use Felo Data API search results in Manus.
- [Instinct](https://openapi.felo.ai/docs/agent-clients/instinct.md): Use Felo Data API search results in Instinct.
- [Claude Duet](https://openapi.felo.ai/docs/agent-clients/claude-duet.md): Use Felo Data API search results in Claude Duet.
- [Other AI Tools](https://openapi.felo.ai/docs/agent-clients/other-tools.md): Connect any client that supports OpenAI-compatible or Anthropic-compatible endpoints.
- [Felo Search API](https://openapi.felo.ai/docs/api-reference/chat.md): Integrate Felo Search API with query analysis, web resources, and citations.
- [PPT Task API](https://openapi.felo.ai/docs/api-reference/ppt-tasks.md): Create PPT generation tasks, list themes, and poll shared task endpoints.
- [Landing Page Task API](https://openapi.felo.ai/docs/api-reference/landing-page.md): Generate landing pages asynchronously and query task history.
- [SuperAgent API](https://openapi.felo.ai/docs/api-reference/superagent.md): Create conversations, consume SSE streams, follow up, and query conversation metadata.
- [LiveDoc API](https://openapi.felo.ai/docs/api-reference/livedocs.md): Create knowledge bases, manage resources, and run semantic retrieval and routing.
- [Web Fetch API](https://openapi.felo.ai/docs/api-reference/web-fetch.md): Extract webpage content as HTML, text, or Markdown with selector and readability options.
- [YouTube Subtitling API](https://openapi.felo.ai/docs/api-reference/youtube-subtitling.md): Fetch YouTube subtitles with optional language and timestamp parameters.
- [X Search API](https://openapi.felo.ai/docs/api-reference/x-search.md): Search X users, tweets, timelines, and replies with pagination metadata.
- [Mindmap API](https://openapi.felo.ai/docs/api-reference/mindmap.md): Create mindmaps from prompts and optional LiveDoc context.
- [Billing and Credits](https://openapi.felo.ai/docs/faq/billing-and-credits.md): How service credits are charged, when to top up, and what a 402 response means.
- [Troubleshooting](https://openapi.felo.ai/docs/faq/troubleshooting.md): Fix authentication, model, streaming, task, and rate-limit problems with a symptom-first checklist.
- Every documentation page also has a plain Markdown version: append `.md` to its URL (for example, https://openapi.felo.ai/docs/getting-started.md). The docs index is https://openapi.felo.ai/docs.md and the changelog is https://openapi.felo.ai/changelog.md; localized variants prefix the locale, for example https://openapi.felo.ai/ja/docs.md.

## API Conventions
- Base URL: https://openapi.felo.ai
- Authentication: Authorization: Bearer YOUR_API_KEY or X-API-Key: YOUR_API_KEY
- Billing: no subscription required. Buy USD-denominated service credits from $1 and pay for actual API usage.