Anthropic rewrites its responsible scaling policy
Version 3.0 took effect, committing the company to publish frontier safety roadmaps and regular risk reports that quantify risk across all deployed models.
Why it matters
A frontier developer committed in writing to periodic quantified reporting on models already in deployment, not only to pre-release capability evaluations.
Version 3.0 is effective from 24 February 2026, and Anthropic describes it as a comprehensive rewrite of the policy. The main changes are publicly released Frontier Safety Roadmaps with detailed safety goals, and Risk Reports that quantify risk across all deployed models. The first such report, the February 2026 Risk Report, is published in redacted form; the next appeared in August 2026 and covers the period since the February one. The policy continued to change through the year: version 3.1 effective 2 April 2026, 3.2 on 29 April, 3.3 on 26 May and 3.4 on 8 July. Earlier versions were 2.2 of 14 May 2025, 2.1 of 31 March 2025, 2.0 of 15 October 2024 and 1.0 of 19 September 2023.