Anthropic, OpenAI to slow frontier pace, let outside auditors in
Amodei's essay embeds outside evaluators in Anthropic and asks Washington for power to block launches; Altman agrees. What it does to your cadence.
Copy markdown
Anthropic embeds METR-style auditors inside
Amodei's first concrete step: give third-party evaluators permanent, employee-level access — badges, desks, and visibility into training pipelines, not just finished models — and let them publish findings without Anthropic's sign-off. It's the first frontier lab to put outside auditors behind the curtain before a model ships.
Altman says OpenAI will do the same
Sam Altman — who called for pacing first — agreed the industry needs to slow frontier development and said OpenAI will also open up to external evaluators. Two of the three labs you build on now signal a pre-release audit step is coming.
'Pacing' is a delay, not a freeze
Amodei is explicit that pacing 'does not mean halting model training or technical progress' — it means taking more time to align each model before pushing capabilities. Read: the gap between a model finishing and hitting your API stretches; the pipeline doesn't stop.
The ask to Washington: power to block releases
The essay's steps two and three call for national laws mandating frontier testing and authorities empowered to block unsafe launches — the same 'duty of care' a Senate bill is already drafting. If it lands, a government gate could sit between a finished model and your access.
It already cost you weeks: Astra slipped
This isn't hypothetical — OpenAI held GPT-6 Astra over cybersecurity concerns before its Sept 4 launch. That's exactly the pacing both CEOs now want to formalize, and a preview of the slower, more staggered cadence to plan roadmaps around.
The counter-take: regulatory capture?
Critics warn that routing releases through incumbent-blessed testing could entrench the biggest labs and squeeze the open-weight models many builders self-host. YC's Garry Tan is pushing the other way — US open-weight labs should get to distill frontier models, treating capable AI as a 'public good.'