Anthropic clarifies RSP v3.1 and advances its Frontier Safety Roadmap

Anthropic’s April 2 update to its Responsible Scaling Policy may look minor on paper, but it matters because RSP language increasingly serves as a public signal of how frontier labs interpret risk thresholds. The company says version 3.1 does not substantially change the policy’s substance, yet it tightens two areas that outsiders were already scrutinizing: how Anthropic defines its AI R&D capability threshold and how much discretion it retains to slow or pause development.

The first clarification addresses a potential ambiguity in version 3.0. Anthropic says language about AI “doubling the rate of progress” could have been read in two very different ways: either as doubling aggregate AI progress or as doubling individual researcher productivity. In version 3.1, the company says it means the former. That is a meaningful distinction because the trigger determines when stronger safeguards or deeper review might be warranted.

The second clarification is governance-oriented. Anthropic now states more clearly that even when its RSP does not explicitly require a specific action, the company remains free to take stronger measures, including pausing development of AI systems if it believes the situation calls for it. In practice, this is a statement about management discretion under uncertainty. It suggests Anthropic does not want outside readers treating threshold rules as a ceiling on caution.

The update also ties into the company’s broader Frontier Safety Roadmap. Anthropic says it has already launched the planned moonshot R&D projects listed in that roadmap and has replaced the original goal of launching them with more detailed goals for ongoing work. It also says it completed a separate goal related to an internal report on how safeguards could be improved through updated data-retention policies. Those details are procedural, but they matter because they show the roadmap is being used as an operational document rather than only as a public promise.

Anthropic also notes that RSP is a living document and expects further revisions as it learns from real-world operation. That is consistent with the rest of the company’s recent safety governance posture, including a March 24 update to its noncompliance reporting and anti-retaliation policy and the February 24 release of RSP version 3.0 with Frontier Safety Roadmaps and Risk Reports.

The broader significance is that frontier labs are now competing not only on model capability and product reach, but also on how legible their safety triggers are to governments, enterprise buyers, and researchers. By clarifying its threshold definitions and discretion to pause, Anthropic is trying to reduce interpretive ambiguity before future model releases push those thresholds closer to practice instead of theory.