haiku 5.5 dropped yesterday and by the evening it was running @moteefai's chat

haiku 5.5 dropped yesterday and by the evening it was running @moteefai's chat we were on gpt-6 luna. anthropic's own numbers had haiku ahead on tool use and reading images, which is basically all our chat does, so I swapped it in with luna as the backup then I had codex test it in the real app. 7 messages, web search, an image upload and an image generation. the whole run cost about 3 cents on anthropic. the one image cost more than every chat turn combined the bigger surprise was speed. luna turns were taking 35 to 65 seconds. haiku did 7 to 16 caching is doing most of the work. 97% of the input on a follow-up came straight from cache, so a normal turn costs about a tenth of a cent