Anthropic's Responsible Scaling Policy 3.0: What It Actually Changes

Anthropic released Responsible Scaling Policy 3.0 on February 24, 2026 — an update to the safety framework the company first published in 2023 that governs when and how it will scale, restrict, or pause development of increasingly capable models. Months on, it's worth revisiting what RSP 3.0 actually changed, since some of the discussion around it at the time overstated its technical scope.

What the Framework Actually Does

RSP 3.0 amends the boundaries between Anthropic's defined AI Safety Levels and updates the company's commitments around pausing training or deployment if a model crosses certain capability or risk thresholds. It is, at its core, a governance and commitment document — it sets out decision rules for Anthropic itself, not a piece of software that auditors log into. It does not introduce any new "interpretability layer" that streams a model's reasoning to outside auditors in real time, and there is no released "Claude 4" model that the policy applies to; RSP 3.0 governs Anthropic's current and future model lineup in general terms.

This is a meaningful but incremental step rather than a technical breakthrough. Independent researchers who study AI governance frameworks have generally described RSP 3.0 as a tightening of commitments Anthropic already had in place, rather than the introduction of new auditing capability. Helen Toner, an independent AI policy researcher (not an Anthropic employee or consultant), has written and spoken publicly about the broader trend of AI labs formalizing scaling policies, a trend RSP 3.0 fits into.

The Interpretability Research Behind the Scenes

Separately from RSP 3.0 itself, Anthropic continues to publish "mechanistic interpretability" research — work that maps specific model behaviors to identifiable circuits within transformer architectures. That research program is real and ongoing, but it is a distinct research effort from the RSP, and it has not (as of this writing) shipped as a live auditing tool that regulators can use to inspect individual model responses.

Industry and Regulatory Response

Policy observers in the EU and UK have pointed to RSP 3.0 as one example of a major AI lab formalizing its internal scaling commitments, and some have suggested frameworks like it could inform future sector-wide requirements. Other labs, including OpenAI and Meta, maintain their own internal safety frameworks and have not adopted Anthropic's specific structure.

What It Means Going Forward

Critics have raised a fair concern: formal scaling policies like RSP 3.0, if they become an informal industry benchmark, could create compliance burdens that are easier for well-resourced labs like Anthropic to absorb than for smaller competitors. That's a governance-cost argument, separate from any claim about new audit technology — and it's the more grounded critique to focus on.