Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war
Source
simonwillison.net
Date
Key takeaways · AI-distilled
GPT-6 Luna costs $0.10/M input and $0.50/M output. Willison notes only GPT-4.1 Nano and GPT-5 Nano, both far weaker, were cheaper among OpenAI releases. GPT-5.6 Terra now costs the same as GPT-6 Sol on input and more on output, which he says removes any reason to pick it.
Opus 5.5 cuts Anthropic's long-standing $5/$25 Opus pricing by 20% to $4/$20, and cache reads fall 60%. Willison says this matters most for long agentic sessions, where 90%+ of input tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → are billed at cached rates.
On Willison's pelican SVG test, Opus 5.5 at max thinking hit the 128,000-token output limit while still reasoning, twice, each run costing $2.56 and taking nearly 20 minutes. He concludes max is effectively useless, while Fable 5.1 at max finished fine.
Willison says his defaults are now GPT-6 Sol in Codex and Opus 5.5 in Claude Code. He moved the Datasette AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → demo to GPT-6 Luna, which he found fast and competent at SQL and at building HTML and JavaScript apps.
Anthropic says Sonnet 5.5 and Haiku 5.5 are coming. Willison asks whether Haiku can compete on price, since Haiku 4.5 costs $1/$5 while GPT-6 Luna is a tenth of that.
Terms in this piece · Glossary
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
Why it matters
GPT-6 Sol and Luna launched at half the price of their GPT-5.6 predecessors, intensifying a price war with Claude Opus 5.5 and Grok 4.7; the piece lays out exact per-token pricing across all of them for engineers picking a model.
