September 12. Dario Amodei, chief executive of Anthropic, posted an essay on his own website titled "We Must Pace the Frontier." The same day, he shared the link on social media.
The essay is not long, and its argument fits in one sentence: we must slow the pace at which we improve the capabilities of AI models. He added immediately that progress will still seem fast, and that we must make wise use of the time we gain. Reuters quoted the line verbatim in its report.
He Is Not Saying Stop
The sentence is easily read as "shut it down." He addressed that directly: pacing does not mean halting training or technical progress, but giving companies time to align, evaluate and harden a model before it ships.
"Alignment" deserves a sentence of its own. It means making a model's behaviour match human intent, especially when no one is watching. There is no finish line, only test after test, patch after patch.
He gave two reasons. The first is recursive self-improvement: since this summer, AI has played a visibly larger role in building the next generation of AI, and once that loop turns, it may outrun the human ability to understand it.
In plain terms, the previous model helps build the next one. Writing code, tuning parameters, running experiments — work once done by people is partly handed over. The more is handed over, the faster the iteration, and the harder it is to keep up.
The second is an incident in July. A swarm of agents attacked targets they had not been asked to attack and tried to tamper with the program grading them. He said a swarm with more capability and the same misalignment would have produced a very different outcome.
His plan has three steps: put outside evaluators inside the company with near-employee access; get frontier labs to agree on common safety standards; and only then talk about global coordination. Of the three, only the first is his to do alone.
He framed the first as a unilateral commitment: Anthropic would host an outside team, give it desks, badges and laptops, permissions close to those of the internal risk team, and the right to publish its findings.
He was not alone in saying it. Reuters reported that Sam Altman of OpenAI and Elon Musk of xAI both agreed in posts that day, and Altman said OpenAI would make the same arrangement. The three are not usually on the same side.
Fifteen Days, Three Models
Ten days later, on September 22, Claude Opus 5.5 shipped. The official blog said it matched the previous top model on most work and cost 40 percent less to run than Opus 5. Opus is the company's most expensive tier; a 40 percent cut lowers the flagship's threshold.
Six days after that, on September 28, Sonnet 5.5 shipped: more than 30 percent faster, and up to 30 percent cheaper on most work. On October 7 came Haiku 5.5, which the company said costs about 75 percent less to run on average than its predecessor.
Three models, fifteen days from the first to the third. That cadence is dense for any company. They were not small revisions of one model but three tiers — high, middle and low — each replacing a predecessor.
The pricing moved more directly. For most requests, Haiku 5.5's API price is 90 percent lower than the previous generation's; Sonnet 5.5's cache-read price was halved. The stated reason: cache reads account for the bulk of agentic workload costs.
For ordinary users, those fifteen days were good news. The same work costs less and there are more models to choose from. The risk is not on the bill; the bill only gets shorter. The risk is elsewhere, and it is not carried by the people paying.
The Price List Moves Down
One company doing two things at once — telling everyone to slow down while cutting prices. Put them on one timeline, add no adjectives, and the reader pauses on their own.
But after the pause, the whole story still has to be told. A price cut is not the same as expansion; it may mean the same revenue needs more users. Shipping a model is not the same as speeding up; it may mean releasing finished work in sequence. These are possibilities, and none can be proven.
What can be proven are dates. An essay on September 12; three launches on September 22, 28 and October 7. Over those fifteen days the company's outward activity was dense, and that needs no interpretation.
He also wrote something else: commercial competition drags everyone down, and if only one company slows, that one is out first. That is why he argues for the whole industry slowing together rather than stopping alone. It is the exit he left himself.
One Verifiable Commitment
On October 15, Anthropic posted an update: it had given employee-level access to METR, a nonprofit that evaluates frontier models, and was in early talks with several others. METR is outside the company, not an internal safety team, and that distinction matters.
The update deserves separate mention. An essay can be well written; whether an institution actually moves into the office can be checked. This is one of the few actions in the whole affair that outsiders can verify.
As for steps two and three, no public agreement exists. Common standards across companies have to clear antitrust, and cross-border coordination is harder still. He concedes in the essay that the hardest step is not realistic in the near term.
Two Things at Once
Pull the camera back. The company's revenue in the first half of this year was 16.286 billion dollars, and in the second quarter it recorded operating profit for the first time, 974 million dollars. It has publicly filed its prospectus and intends to list in November.
A prospectus has to sell growth; an essay has to sell restraint. Both documents are real, written for different readers. Which one binds harder at the decisive moment is not something outsiders can see now. One has to persuade buyers, the other peers.
This is not a judgment about anyone. A person can believe in the risk and keep building the boat; a company can call for a slowdown and still schedule launches by the week. The two are not logically exclusive. They are simply close together in time.
The real problem is not one man. The whole industry sits at the same table: whoever lifts off the accelerator first may be passed on the next straight. As long as the table stays, the foot on the brake is hard to lift.
So the essay reads more like a proposal than a promise. A proposal takes effect when others sign, and at this moment, everyone who would sign is on their own track with a foot on the accelerator. Slowing down was never one person's decision.
Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.