Just did a deep dive on the current text-to-image landscape (Sep 2026). Benchmarked 15 of the top models — paid API + open-weight/local — across two independent human-preference ar

Just did a deep dive on the current text-to-image landscape (Sep 2026). Benchmarked 15 of the top models — paid API + open-weight/local — across two independent human-preference arenas (Artificial Analysis Image Arena + LMArena), plus price and speed. Highlights: 🥇 GPT Image 2.5 Sunburst is #1 in both arenas (1195/1424) — but 109s per image and $0.21 💸 Best value: Meta Muse Image — top-5 quality at just $0.01/image ⚡ Fastest: Nano Banana 2 (9s) for paid, Z-Image Turbo (2s, free locally) for open 🏠 Best open-weights: Qwen-Image-2.1 leads (1035 Elo). Sleeper pick: Ming-Image-0.1-Design (MIT) — built for design, UI, text rendering & transparent backgrounds The model breakdown: 🥇 GPT Image 2.5 Sunburst (OpenAI, Sep 2026) — 1195 AA / 1424 LMArena, 77% win rate. The new quality ceiling, but it's the slowest and priciest in the table: ~109s per image, $0.21. 💸 Muse Image (Meta, Jul 2026) — 1111 AA / 1276 LMArena at $0.01/image, ~14s. Meta's agentic model (reasons before rendering) hits top-5 quality for ~1/20th the price of the leader. ⚡ Nano Banana 2 (Google, Feb 2026) — Gemini 3.1 Flash Image: 1124 AA / 1261 LMArena in ~9s. The fastest model that stays in the top tier. ⚡ Z-Image T