Ten days after the pacing call, the price fell

On 12 September the two largest AI laboratories in the United States agreed in public that the frontier was moving too quickly. On 22 September they both shipped, and both shipped cheaper.

Anthropic released Claude Opus 5.5 on Tuesday. It is priced at $4 per million input tokens and $20 per million output tokens, against $5 and $25 for Opus 5, which it replaces at the top of the Opus line. Cache reads fall further, from $0.50 to $0.20 per million, a 60 per cent cut, on the company's published price list. The batch rate drops from $2.50 and $12.50 to $2 and $10. The premium fast mode falls from $10 and $50 to $8 and $40. Anthropic says the combined effect is roughly 40 per cent less for a typical workload, and that output is generated more than 30 per cent faster than Opus 5.

OpenAI answered within about an hour with GPT-6 Sol and GPT-6 Luna, a mid-tier and a budget model beneath the GPT-6 Astra flagship released earlier this month. Sol is reported at $2 per million input tokens and $10 per million output; Luna at 10 cents and 50 cents. Both are described as roughly half the price of the GPT-5.6 models that carried the same names. Anthropic and OpenAI released the two within hours of each other, SiliconANGLE reported.

The capability claims are the vendors' own and should be read as such. Anthropic puts Opus 5.5 at 66.4 per cent on Terminal-Bench 4.0 against 52.3 per cent for Opus 5, at 54.4 per cent on FrontierCode v1.1 against GPT-6 Astra's 53.3 per cent, and at 1846 Elo on GDPval-AA v2.1 against 1735 for its own larger Fable 5.1 model. The pitch is Fable-class work at Opus prices. Nobody outside the labs has yet run these numbers independently.

The explanations offered are unremarkable and probably accurate. OpenAI attributed its reductions to improvements in caching and inference and said it was passing the savings to customers. Anthropic used almost the same words. Inference costs fall as models are distilled, quantised and served on better hardware, and in a market with two credible suppliers the savings do not stay with the supplier for long. "OpenAI and Anthropic are engaged in a price war that is driving down the price of AI," Ara Kharazian, lead economist at the spend-management company Ramp, told Fortune. He noted that the bull case for these companies assumed the opposite: that better models would command higher prices.

What makes the timing consequential is the legal position. Ten days before these releases, Dario Amodei published an essay calling for the industry to pace the frontier so that safety practices could keep up with capability, and Sam Altman, Elon Musk and Demis Hassabis endorsed it the same day. Washington and Beijing both rejected the idea within 48 hours. Four subscribers then sued the four companies in federal court in San Francisco, alleging that an agreement among rivals to withhold improvement is a restraint of trade under section 1 of the Sherman Act, and seeking treble damages.

The defendants now have an obvious exhibit. Within ten days of the essay that forms the centrepiece of the complaint, two of them cut prices by 20 to 60 per cent and shipped models they claim are materially more capable. A cartel that agreed to restrain output does not usually behave like that in the fortnight after it forms.

The argument is strong but it is not dispositive, and it is worth being precise about why. The complaint alleges a restraint on the rate of capability improvement, not on price. Cutting the price of a given level of capability is consistent with an agreement not to advance capability as fast as one otherwise would — indeed it is what a firm does when it cannot compete on the other axis. A plaintiff will also point out that a model released on 22 September was trained and evaluated long before 12 September, so the release says nothing about what either company chose not to build. And section 1 cases turn on whether an agreement existed, not on whether it was successfully carried out.

Against that, the labs have a straightforward account of their own conduct. Anthropic says Opus 5.5 was reviewed before release by external evaluators including METR, which is the practice Amodei proposed in the essay, and frames the pacing argument as being about safeguards in high-risk domains rather than about withholding products. The pledge as its authors describe it was never a promise to stop shipping. It was a promise to let someone else check first.

For buyers, the antitrust question is secondary to the arithmetic. The cost of the most capable tier of model has fallen by about a fifth in two months, and the cost of cached context by more than half. That changes which applications clear their hurdle rate, and it changes them in the direction of more usage, not less. It is a strange outcome for a fortnight that began with both suppliers agreeing, in public, that things were moving too fast.