Anthropic Releases Responsible Scaling Policy Version 3.0 With New Operating Model for ASL Thresholds

What changed in Anthropic's policy framework

Anthropic published Responsible Scaling Policy (RSP) Version 3.0 on February 24, 2026, positioning the document as an operational update rather than a branding exercise. RSP is Anthropic's voluntary framework for reducing catastrophic risks from advanced AI systems. The company first introduced the policy in September 2023 and has treated it as a living document tied to model deployment decisions.

The core architecture remains familiar: conditional "if-then" commitments connected to AI Safety Levels (ASLs). If model capabilities cross specified thresholds, stronger safeguards are required. Anthropic says this structure has had real internal force. In the post, the company points to its activation of ASL-3 protections in May 2025 and ongoing work to improve safeguards such as constitutional classifiers and other controls against misuse.

Why Version 3.0 exists

The major shift in Version 3.0 is how Anthropic handles ambiguity at the frontier. The company argues that some high-stakes capability thresholds are no longer clean pass/fail events. In areas such as biological risk, fast tests can suggest elevated concern, but they may not provide definitive evidence for how close systems are to severe real-world misuse. Anthropic references additional evidence gathering, including wet-lab related research, but notes that evaluation cycles can lag model progress.

That gap creates a policy problem: thresholds are still useful, but rigid trigger logic can become brittle when measurement is uncertain and the external policy environment moves slowly. Anthropic says Version 3.0 addresses this by separating what can be achieved unilaterally now from what likely requires broader coordination across industry and government. Instead of over-promising at higher ASL tiers, the company introduces publicly declared targets and commits to grading its own progress in public.

Why this matters for the broader AI ecosystem

• It reframes frontier safety policy as a continuous operating discipline, not a one-time publication.
• It acknowledges evaluation uncertainty as a first-order governance issue, especially for catastrophic-risk domains.
• It strengthens transparency as a practical accountability tool when formal regulation is still catching up.

For operators, researchers, and policymakers, RSP Version 3.0 is significant because it documents the tradeoff many labs now face: maintain strict safety intent while adapting implementation to evidence quality, deployment tempo, and real-world governance constraints.