OpenAI and Anthropic Agree AI Should Slow Down — While They Keep Shipping
Dario Amodei of Anthropic and Sam Altman of OpenAI agree that AI development needs to slow down.
That agreement was reported on September 15, 2026 Associated Press. It puts the heads of two frontier labs on the same side of a capabilities question, even as both labs continue to ship models, APIs and enterprise products.
Amodei made the more explicit public case. He called on AI companies to "slow the rate at which they advance model capabilities" Reuters. He then wrote an essay calling for slowing down AI development, according to September 14 reporting. He also urged regulations requiring "independent audits" for the AI industry.
The political response was split. U.S. President Trump dismissed AI safety alarms in September 2026, saying the U.S. "already has tools to police the AI industry" Reuters. A China state newspaper criticized Anthropic's calls to slow AI development as a "Cold War tactic" Reuters.
OpenAI has its own safety language on the record. The company publishes an official page titled 'An Alien Mind' on its own domain. That page states the time calls for "extreme caution" and that "no one is prepared for the consequences of a continued rapid rise in machines."
OpenAI also publishes an official page titled 'AI progress and recommendations.' That page, dated November 6, 2025, states OpenAI is "deeply committed to safety."
Those statements sit alongside a series of product announcements. On September 16, 2026, OpenAI published 'Reimagining advertising with AI,' 'How to connect AI usage to business value' and 'Our framework for reporting model misalignment.'
The prior week brought further releases. On September 11, OpenAI published 'Rapidly scaling online storage to serve over 1 billion ChatGPT users.' On September 10, it published 'How a researcher uses Codex and ChatGPT to search for new antimicrobial molecules,' 'Now everyone can put data to work,' 'Introducing ChatGPT for Financial Services,' 'Build more natural voice experiences with GPT-Live-1 in the API' and 'Introducing the Agents API.'
Anthropic is also shipping. On September 1, 2026, it announced Claude Fable 5.1 and Claude Mythos 5.1. The company describes the pair as its "most advanced models for coding and knowledge work."
The broader context here is the difference between model risk and deployment risk. Think of it as improving the engine versus improving the car around it. Slowing capabilities means slowing pre-training scale, which is training with more data and computing power, along with new reasoning behavior for multi-step problems, greater autonomy to act without close supervision, and gaps in testing. Product work is different. It covers packaging, retrieval from outside sources, voice latency or response delay, coordination of agent tools, storage, and meeting rules for specific industries. The September publications from OpenAI fall largely in that second group. Amodei's audit proposal addresses the first risk, with third-party access, repeatable evaluations and disclosure duties across labs. Trump's response holds that existing oversight is enough. Beijing's response treats the slowdown call as geopolitical rather than technical.
In my view, worth flagging is how closely the slowdown language tracks with scale. A lab serving over 1 billion ChatGPT users and supporting research agents for antimicrobial discovery is no longer discussing hypothetical misuse. Misalignment reporting, independent audits and storage infrastructure are daily operational questions at that size. My kids grew up with search, then smartphones, then generative chat. Each shift looked gradual until schools, workplaces and support systems had to adjust quickly. Enterprise AI feels similar now.
The reason this moment is still constructive is that the unglamorous work makes routine use possible. Reporting frameworks for model misalignment, audit requirements, clearer measurement of business value and specific controls for financial services let powerful systems be used with care. If labs can agree on restraint for capabilities while competing on safety tooling, reliability and verifiable deployment, technology professionals gain something usable. Caution, in that sense, enables adoption rather than blocking it.
