Anthropic said slow down. Ten days later it shipped its fastest, cheapest model yet.

Here is the sequence, in order. On 12 September, Anthropic's chief executive publishes an essay telling the entire AI industry it needs to deliberately slow the pace at which it improves model capabilities, because safety work can't keep up otherwise. Ten days later, on 22 September, his own company ships a model that is faster, cheaper and — by its own account — more capable than anything it has released before. Both things can be true at once. It's still worth sitting with the contradiction before accepting Anthropic's explanation of it.

Start with what you don't have to take on faith. The pricing is public and falsifiable: Opus 5.5 is $4 input and $20 output per million tokens, against Opus 5's $5/$25 — a real 20% cut on the headline rate, and a real 60% cut on cache reads, from $0.50 to $0.20 per million. Cache reads are where the money actually is in sustained agentic work, so that number matters more than the flashier '40% cheaper to run' framing, which bundles token pricing with unstated assumptions about how many tokens a task now takes.

Every score in the launch — 66.4% on Terminal-Bench 4.0, 1846 Elo on GDPval-AA v2.1, 81.8% on OSWorld 2.0 — was run by Anthropic, on Anthropic's chosen effort settings, in Anthropic's chosen harness. The company's own footnotes are unusually candid about this: it notes its Terminal-Bench 4.0 reproduction of Opus 5's score (52.3%) sits within the standard error of the public leaderboard figure (51.8%), and that GPT-6 Astra's numbers are 'as reported by OpenAI' — a polite way of saying nobody has actually run all these models side by side under one set of rules. That's normal for a launch post. It is not the same thing as an independent result, and the coverage repeating these numbers as fact — rather than as 'Anthropic says' — is doing readers a disservice.

The more interesting admission is buried in the prose rather than the tables: Anthropic writes that 'the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest' in its own internal use. Fable 5.1 costs $10 input / $50 output per million tokens — two and a half times Opus 5.5's price. If Anthropic's own staff can't reliably tell the two models apart on real work, that is a pricing problem for Fable, not a compliment to Opus 5.5. Nobody at Anthropic phrased it that way, for the obvious reason that Fable is still their premium product.

Anthropic says Opus 5.5 was tested pre-release by Frontier Design and METR — real external evaluators, genuinely independent of Anthropic — and that it is 'the strongest-performing model we've tested to date' on Anthropic's automated behavioural audit, which is Anthropic's own test, scored by Anthropic. It also says Opus 5.5 ties Fable 5.1 for the lowest prompt-injection success rate on a benchmark run by Gray Swan, an AI security firm. That's a genuine third party — but Anthropic doesn't say whether Gray Swan's evaluation was commissioned by Anthropic, run independently, or something in between, and the launch page doesn't link the underlying report. METR and Frontier Design get named; the depth of what they actually tested does not, beyond a pointer to Anthropic's own system card.

That gap made a better headline than 'lab revises its own pricing tiers', which is presumably why most coverage led with it. The more telling fact sits one line down: OpenAI's new models are priced at roughly half of the GPT-5.6 generation, which OpenAI attributes to caching and inference improvements rather than a deliberate loss-leading price war. xAI's Grok 4.7, released the day before at unchanged $2/$6 pricing, chose the opposite move — hold price, improve the model. Three different labs, three different pricing philosophies, inside 48 hours. That is the actual competitive story, and it has nothing to do with who hit publish first.