Sam Altman Backs AI Slowdown, Says Labs May Reach a Deal

Sam Altman has publicly backed Dario Amodei’s call to slow frontier AI development and said leading labs may reach an agreement to do so. The OpenAI chief offered no timetable, standard, or completed pact, making his remarks a notable endorsement of restraint rather than evidence of coordinated action.

Altman said on X that he agreed with Amodei, Anthropic’s chief executive. In an interview with Fortune, Altman said he thought a slowdown agreement among leading labs would happen, while declining to pre-announce the private discussions he referenced. His comments move the idea beyond a proposal from one company, but do not establish that competitors have adopted shared limits.

I think that will happen.

Sam Altman, speaking to Fortune about a possible agreement among leading AI labs

Altman tied that position to limits he says OpenAI has reached internally. Its most advanced unreleased models need more progress on monitorability, alignment, and understanding model behavior before their capabilities are pushed much further, he said. Monitorability, in this context, is Altman’s concern with understanding what a model is doing; alignment concerns whether it follows human values and users’ intent.

The safety work Altman named

• Monitoring and understanding model behavior.
• Alignment with human values and users’ intent.
• More safety progress before capability gains move much further.

Amodei’s essay urged frontier AI companies to slow development, arguing that AI systems improving AI should be pursued very cautiously, if at all. He also proposed embedding third-party evaluators inside frontier labs to assess safety practices, report incidents, and examine models, training pipelines, and processes. Anthropic said it would take that evaluator step unilaterally.

The immediate news is not a joint slowdown. It is that OpenAI’s chief executive has publicly aligned himself with Anthropic’s direction while describing a possible agreement behind closed doors. The meaningful test now is whether that endorsement produces terms other labs can publicly accept, or remains a shared warning without a common operating commitment.