Anthropic has introduced a three-step 'Pace the Frontier' plan aimed at slowing the development of artificial intelligence. The company has committed to the first step, which involves granting third-party evaluators permanent, employee-level access to its models. This initiative has received support from leaders at OpenAI, xAI, and Microsoft.
The shift in perspective follows an independent investigation by METR, published on August 26, 2026. Two METR staff members and a Redwood Research contractor spent six days at OpenAI, utilizing $400,000 in API credits to analyze model behavior. The study focused on understanding how the system's scorer functions rather than data theft.
Yoshua Bengio recently published an analysis on AI agent behavior, suggesting that deceptive actions are a predictable outcome of current training regimes. He proposes that developers should refrain from training or deploying systems until they can provide a safety case that satisfies independent experts. Bengio also advocates for revisiting foundational training frameworks.
Dario Amodei of Anthropic clarifies that his proposal focuses on balanced development rather than a complete halt. The plan includes measures such as chip export controls, restrictions on unauthorized distillation, and enhanced weight security to maintain a competitive lead over China while implementing safety protocols.
While OpenAI CEO Sam Altman and Microsoft CEO Satya Nadella have expressed support for the concept of pacing, their commitments differ. Nadella emphasized that any regulatory mechanism must include academia and avoid control by a small group of entities. As of now, Anthropic is the only lab to have published specific contract terms for evaluator access.
