Version 3.3 is live. Anthropic’s Responsible Scaling Policy page confirms the update directly: “Last updated May 26, 2026” with “Version 3.3 and redline (effective May 26, 2026)” listed alongside the version history. This isn’t a rumor or a third-party report. It’s on the primary source.
The version history is what makes this interesting. v3.0 took effect February 24. v3.1 followed April 2. v3.2 arrived April 29. v3.3 landed May 26. Three updates in 90 days, with the last two separated by fewer than 30 days each. That’s a substantially faster iteration cycle than RSP’s initial pace.
The policy covers Anthropic’s framework for evaluating when AI capability development requires additional safety measures, specifically around thresholds that, when crossed, trigger mandatory safeguards before further development proceeds. The RSP is the document that governs what Anthropic can and can’t do with models like Mythos.
What changed in v3.3: the public document confirms updates to safety thresholds and references to off-cycle risk update protocols. The specific content of the threshold revisions isn’t fully detailed in the publicly available version, some elements appear to be redacted or addressed in the separately distributed redline. Treating the threshold changes as confirmed specifics isn’t possible from the accessible source; treating the fact of revisions as confirmed is.
Don’t read the pace of updates as instability. Faster RSP iteration is more likely to reflect two things: Anthropic’s capabilities are advancing faster than the original policy cadence anticipated, and the policy framework is being actively stress-tested against new capability evaluations rather than sitting static. A policy that doesn’t update when capabilities change is a policy that’s failing at its job.
The off-cycle protocol refinement matters for practitioners tracking Anthropic’s safety posture. Prior RSP versions established the framework for when Anthropic triggers a safety evaluation outside the regular review schedule. v3.3 reportedly refines how that process works, per the policy’s framing, the specific operational details aren’t confirmed from the public document.
What this means for teams depending on Anthropic’s safety commitments: the RSP’s rapid iteration is a signal that the framework is being actively maintained. That’s good. The corresponding challenge is that each version change can shift the threshold at which Anthropic’s safety commitments apply. Compliance teams and enterprise buyers with Anthropic contracts that reference the RSP should confirm which version their agreement references, and whether the version 3.3 changes affect their specific use case or deployment context.
Who This Affects
The cross-pillar implication is direct. The same RSP that governs Anthropic’s commercial model development also governs the Mythos access structure and the Glasswing coordination program. A threshold revision in v3.3, even one not publicly detailed, can affect which capabilities Anthropic is permitted to deploy and under what conditions. For Glasswing partners and enterprises using Claude in regulated environments, that makes the RSP version history a document worth monitoring on the same cadence Anthropic publishes it.
Watch for two things: first, whether Anthropic publishes a blog post detailing the v3.3 changes in accessible language (prior RSP updates have sometimes been accompanied by explanatory posts). Second, whether the next update arrives within another 30-day window, if the cadence continues at this pace through June, it suggests a capability threshold crossing is being managed in real time.