Anthropic has dropped a central commitment to pause AI training or withhold models when required safeguards were unavailable, revising its voluntary Responsible Scaling Policy to emphasize public safety goals, regular risk reporting and conditional delays. The company announced version 3.0 on February 24, 2026.
The framework, introduced in September 2023, linked increases in model capabilities to stricter safeguards. It required Anthropic to pause training of more powerful models if development outpaced its ability to meet the necessary safety standards.
“We felt that it wouldn’t actually help anyone for us to stop training AI models,” chief science officer Jared Kaplan said. He argued that unilateral commitments made less sense while competitors were “blazing ahead.”
Anthropic said evaluations had proved less conclusive than anticipated about whether models had crossed specified capability thresholds. The company cited biological knowledge tests that, it said, were insufficient to establish either low or high risk. It also argued that some safeguards for more advanced systems could require collective action beyond what a single developer could achieve.
Public safety goals and conditional delays
The revised policy separates safeguards Anthropic plans to pursue itself from more ambitious measures it recommends across the industry. It requires a public Frontier Safety Roadmap covering security, alignment, safeguards and policy. Anthropic describes the roadmap targets as nonbinding goals against which it will publicly assess its progress.
Anthropic says it will publish Risk Reports every three to six months, describing model capabilities, potential threats, active safeguards and its assessment of overall risk.
The revised approach retains conditional delays. Anthropic says it will delay AI development if its leaders consider the company to be leading the field and judge catastrophic risks to be significant. It also commits to matching or surpassing competitors’ safety efforts.
Chris Painter, director of policy at AI evaluation nonprofit METR, reviewed an early draft of the policy. He welcomed its emphasis on transparent risk reporting and public safety roadmaps but expressed concern that moving away from binary thresholds could allow risks to increase gradually without a clear trigger for halting development.
Funding, products and government negotiations
The revision followed Anthropic’s February announcement that it had raised $30 billion at a $380 billion post-money valuation. The company said the investment would support research, product development and infrastructure expansion.
Recent product announcements included Claude Code in Slack, introduced as a beta research preview, and Sonnet 4.6, which Anthropic announced with a 1 million-token context window in beta.
Anthropic was also involved in negotiations with the Pentagon over Claude’s usage limits. CEO Dario Amodei said the company would maintain restrictions on mass domestic surveillance and fully autonomous weapons.
