Anthropic Says AI Is Moving Too Fast. It Wants Outside Checkers Inside the Lab
Anthropic CEO Dario Amodei laid out three ideas to "pace the frontier" in a blog post reported on September 12, 2026, and said Anthropic will do one of them on its own. TechCrunch
Amodei wrote, "We must slow the pace at which we improve the capabilities of AI models." He said two things changed his thinking: the OpenAI-Hugging Face hack and AI getting much better much faster in recent months, especially at helping build the next generation of AI.
The step Anthropic can take alone is to let outside checkers work inside the lab. Amodei proposed "embedded evaluators" from outside groups such as METR to check if Anthropic keeps its pacing and safety promises and to make sure safety problems are reported.
He compared them to government watchdogs who sit with bank workers. The plan would give evaluators badges, desks, laptops, and access mostly like internal risk assessment teams have, with exceptions when required by law or contracts.
The broader context here is why inside access matters. Internal risk teams see the full setup: the systems, the helper software around the model, the test tools, and the logs of problems. Outside reviewers often see much less, only what comes through a limited outside connection, called an API, or a prepared demo. Matching inside access would change safety checks from occasional tests to steady watching.
Coordination among labs and a role for government
The rest of the plan needs others to act. He called on governments to require other frontier companies to adopt the same embedded-evaluator commitment.
He also called for leading AI companies in democratic countries to agree on shared safety standards and limits on the speed of unchecked AI progress. He said the U.S. government should help lead or at least allow those safety talks and give a narrow antitrust waiver for certain kinds of safety conversations.
To understand that legal request, it helps to know the current rule. Rival labs cannot jointly set limits on how fast they build or release AI without risk of breaking collusion law. A waiver limited to safety talks would let them coordinate on safety without allowing wider coordination on products or prices.
The pacing call did not emerge in isolation. Anthropic and OpenAI backed an employee-led initiative asking the U.S. government to "deliberately pace" frontier AI development, with more than 1,200 employees urging the U.S. government to pace AI growth. The Hill Anthropic separately stated it supports a petition signed by its CEO, several co-founders, and senior staff to pace the frontier of AI development so society can prepare.
A fast product cadence alongside a slowdown call
The September 12 message came during a quick series of releases. Anthropic introduced Claude Opus 5 on July 24, 2026. It published "Investigating three real-world incidents in our cybersecurity evaluations" on July 30, 2026. It said Mariano-Florentino (Tino) Cuéllar would join as Chief Global Affairs Officer on August 4, 2026. It opened a research preview of the Model Hardware Standard to a first group of scientific research labs and advanced manufacturers on August 27, 2026. Anthropic
Anthropic introduced Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. It called them its most advanced models for coding and knowledge work. It listed "Developing Enterprise Frontier Safeguards with our customers" dated September 1, 2026. It published "Detecting and countering misuse of AI: September 2026" on September 10, 2026.
The broader context here is that building more capable AI and adding safety tools are happening at the same time. Capability work continues. Safety tools, misuse detection and business controls ship alongside it. The argument here is that the two tracks need clear linkage through outside checking.
In my view, the plan is best read as basic safety plumbing, not a call to stop. Badges, desks and access to logs are boring details. They create clear records. They let an outside party with different incentives see what was tested, what failed, and what was reported. For businesses using these models, that kind of outside presence could eventually make buying, problem response and legal compliance simpler, if governments make the rule apply to all labs.
Worth flagging, working together is the hard part. One lab can open its doors tomorrow. Shared safety standards and shared limits need rivals to agree on how to measure, what the limits are, and how to enforce them. That is why the call for U.S. government help and a narrow antitrust waiver sits next to the step Anthropic can take alone. One can be done now. The other needs government action.
Looking at the longer arc, the hopeful read is simple. Slower growth in what models can do, plus stronger checking, could make the systems easier to use in real work. Testing that keeps up with training helps with speed, reliability and security. Pacing, in that sense, is not only about lowering risk. It gives everyday practices, tools and rules time to catch up with what models can already do.
