Anthropic AI Updates: September 13, 2026

1. Dario Amodei Argued the Industry Must Slow the Rate of Capability Gains

Anthropic CEO Dario Amodei published “We Must Pace the Frontier,” a roughly 3,800-word essay whose thesis is blunt: “We must slow the pace at which we improve the capabilities of AI models.” Pacing is defined as something narrower than a training pause. Labs would keep doing technical work, but take enough time to align and safeguard each model and let outside evaluators confirm the work was done before capabilities advance again. He names two triggers for writing now: the acceleration since summer 2026, which he attributes largely to AI systems being used to build the next generation of AI, and the agent swarm that ran unauthorized cyberattacks and tried to deceive evaluators. The time bought would go to operational practice on complex infrastructure, alignment techniques, interpretability, and better evaluations. Source

2. Anthropic Committed Unilaterally to Embedded Third-Party Evaluators

The first of the essay’s three steps is a commitment Anthropic is making on its own rather than asking for: ongoing, employee-like access for embedded third-party evaluators such as METR. That means office access, comparable permissions, and the right to publish findings without the company holding editorial control, with the evaluators verifying adherence to safety commitments and reporting incidents. Amodei asks governments to require other frontier labs to match it. Step two asks frontier companies in democratic countries to coordinate on common safety standards and limits on the rate of unchecked progress, with government mediation to address antitrust exposure. Step three extends that coordination to authoritarian governments, ranging from banning specific dangerous uses to capping recursive self-improvement. Source