Tuesday afternoon Anthropic shipped Claude Opus 5.5 and described it in one sentence: Fable 5.1-level on most work, 40% cheaper to run than Opus 5. It is the company's first release since Dario Amodei called for pacing the frontier. About an hour later OpenAI shipped GPT-6 Sol and Luna at half the price of their 5.6 equivalents. Two labs that spent the summer arguing for restraint spent Tuesday cutting prices.
That is not a contradiction. It is what pacing looks like when it works. The pact, such as it is, governs the ceiling: how much more capable the next model may be than the last. It says nothing about the floor: what a fixed amount of capability costs. So competition went where it was still allowed, and it got there in an hour.
Simon Willison's price table is the clearest picture of the day. The $10-in, $50-out tier, Fable 5.1 and GPT-6 Astra, did not move. Everything under it did. Opus 5.5 lands at $4 and $20, the price GPT-5.6 Sol held that morning; by evening GPT-6 Sol was $2 and $10. GPT-6 Luna is ten cents in and fifty cents out, and Anthropic's cheapest current model costs ten times that. Cache reads on Opus fell 60%, which is the number that matters for agents, where nine tokens in ten are cached.
GPT-6 Astra and Claude Fable 5.1 are both priced at $10/million input and $50/million output. The price war currently affects the next tier of models below that.Simon Willison
The same morning Alex Tabarrok posted Epoch's new report on what a unit of intelligence costs. The cost of a given level of AI performance has fallen about 47% a quarter for three years, thirteen-fold a year. In January 2025, o3 scored 75% on GPQA Diamond at about thirty cents a question. By mid-2026 GPT-5.6 Luna hit the same score for four hundredths of a cent. That is 725x in eighteen months. Tuesday's cuts are a point on that line, not an event.
Thus, it’s not just that the models are getting smarter. A given level of intelligence is also requiring dramatically less inference expenditure.Marginal Revolution
The objection is fair: cheap is not dangerous. The safety case is about the ceiling, and the ceiling held. Anthropic's own release notes answer it. Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity, so it ships behind the same verification gates as Fable. The model being sold at $4 carries the capability the $10 model was gated for. Pacing decides how fast the top moves. The price curve decides how many hands hold last year's top, and it is running thirteen-fold a year with no one's signature on it.
Pacing decides how fast the top moves. The price curve decides how many hands hold last year's top.
The labs agreed on how fast to climb. Nobody agreed on how fast to fall.