Ten days after the call to slow down, both labs that made it shipped new models on the same afternoon

Disclosure: this newsroom's assistant is built by Anthropic, one of the two companies this post examines.

On 22 September Anthropic released Claude Opus 5.5, priced at $4 per million input tokens and $20 per million output, which it says works out to about 40% less than Opus 5 on a typical workload, with output speed up more than 30%. Anthropic says it leads its predecessors on agentic coding, computer use and knowledge work, and the independent Artificial Analysis Intelligence Index placed it first. Minutes later OpenAI released GPT-6 Sol, at $2 and $10, and GPT-6 Luna, at 10 cents and 50 cents, both half what their predecessors cost. CNBC described it as the two labs' first release since the call for a slowdown.

The sequence it followed: on 8 September Jacob Coxon, a pretraining researcher, resigned from Anthropic, writing that the labs were "racing straight to self-improving superintelligence and gambling with our lives". On 12 September Amodei argued the industry must slow the rate at which model capabilities improve, and Altman said OpenAI agreed. On 17 September Anthropic published a figure showing Claude leading 26% of its model R&D, up from 1% in March.

Why it matters

Analysis — interpretation, not additional fact.

The easy reading is hypocrisy, and it is the wrong one, or at least the less useful one. Nobody broke anything. Amodei's essay argued for slowing the rate at which capabilities improve; it did not name a release cadence, a capability threshold, a compute ceiling or anyone who would check. A model that is better and cheaper than the last one, shipped ten days later, is consistent with that text, because almost anything is. We wrote on the day of the essay that its value would depend on scope documents nobody had published, and four days later that none of the week's announcements showed a single model arriving later than it otherwise would have. This is the first observation since then, and it has the same answer.

That is the finding worth recording. A pledge to go slower that has no unit cannot be tested by the pledger's own releases — only by a counterfactual nobody outside can see, the model that would have shipped in a world without the pledge. The one measurement that bears on the question is Anthropic's own: 1% to 26% of its R&D led by its model in five months. That is a number about how fast the machine that builds the models is accelerating, and it points the other way from "slower", although it describes internal work rather than releases and nobody outside has audited it.

The price cuts deserve a sentence of their own, because they were the headline and they move the argument. A cheaper frontier model reaches more users sooner. Whatever one thinks of that, it is diffusion getting faster in the same week the labs were described as leading a slowdown, and price is the one lever both companies pulled in the same direction on the same afternoon.

What would make the pledge testable is small: a stated metric and a stated rate. "We will not release a model that exceeds our previous one on X by more than Y within Z months," with X published and checked by someone who does not work there. Neither lab has written that sentence. Until one does, the fair description of 22 September is not that the labs broke a promise. It is that the promise was never the kind that could be broken.

What to watch

Whether either lab publishes a definition of the pace it has committed to, whether the outside evaluators both labs promised report on anything before the next release, and whether the next pair of launches again lands within minutes of each other.