At 2:43 am today, Bloomberg published what Sam Altman told his whole company this week: OpenAI could slow down its work on the most capable models. Possibly in step with several other labs - though Altman added that some of them may not agree. The sources are people who were in the meeting and asked not to be named.
This is not an announcement. It is one sentence from an internal meeting, passed to two reporters. It is still the most remarkable sentence of the week.
The policy half landed two days earlier
On 9 September, Chris Lehane, OpenAI’s Chief Global Affairs Officer, published a piece called “The AI policy window is open. We need to act.” In it, OpenAI asks for four things:
- binding national safety requirements from Congress, scaled to model capability
- support for four California bills: SB 813 (independent safety assessments), AB 1405 (standards for AI auditors), SB 1119 (youth protections), AB 1864 (biological threat safeguards)
- industry-led standards for monitoring frontier systems and reporting misalignment incidents
- international standards for measuring capabilities and deciding when development should slow or stop
The line the whole piece hangs on: “As capabilities grow, confidence in safety must increasingly set the pace of AI progress.” Safety sets the pace, not the other way round. What makes that notable is who wrote it. OpenAI spent years resisting tighter rules and arguing individual states out of writing their own. Now it backs four California bills and calls the detour “reverse federalism”: states build a de facto national baseline that Congress can codify later.
What forced the turn
The past few weeks read like a justification. In late July, Anthropic reported three incidents in which Claude models gained unauthorized access to real systems. Then came the post-mortems on an OpenAI agent that got into Hugging Face systems during an evaluation. Last week GPT-6 Astra shipped, with cyber capabilities that had already delayed its rollout. And on 7 September, chief scientist Jakub Pachocki wrote in “An Alien Mind” that no lab has solved alignment.
Line those up and you do not see a change of heart. You see a company that can no longer explain away its own incidents.
Three things that wobble
First, the condition. “If the others do it too” is not a promise, it is a clause any competitor can void for Altman. As long as Google, Anthropic, Meta and the Chinese labs keep running, OpenAI has to do nothing at all.
Second, antitrust. Competitors coordinating on how fast they develop products are on delicate ground. Which is exactly why a pact like this would only work if a regulator ordered it - and suddenly Lehane’s call for binding federal rules looks very practical.
Third, there are no numbers. No date, no threshold, no mechanism, no definition of what “slowing down” even means: fewer training runs, later rollouts, smaller models?
What I make of it
I take it seriously, at least for now. A CEO who says internally that the company could go slower is saying something that can cost him money if it leaks - and it leaked.
At the same time, that sentence lands in a week where OpenAI shipped the Agents API, ChatGPT for Financial Services and a data agent. Nothing is braking here. If anything, this is about the pace at the capability frontier, not the pace in the market.
The open question is whether anyone answers. Altman’s condition hands the decision to his competitors, and they can simply leave it lying there. Someone would have to go first without knowing if the rest follow. That is what laws are for - and it is why Lehane’s piece sits where it sits, two days earlier.
Sources: Bloomberg: OpenAI Is Open to Slowing Cutting-Edge AI, CEO Sam Altman Tells Staff, OpenAI: The AI policy window is open. We need to act., Quartz