Red teams access Niptune v3 in lead-up to new Claude model

What we know so far: Claude Neptune v3 is the next version of Anthropic's safety system, which is typically a predecessor of a new model upgrade. Rumours suggest that we may see Claude 4.1 or 4.2 released within the coming weeks.

· 1 min read
Claude

Anthropic is currently testing a new iteration of its safety system, Niptune, with red teams gaining access to Claude Niptune v3. Niptune first appeared just before Claude 4 as a jailbreak prevention system using constitutional classifiers, and it seems the company continues to use this staged upgrade approach. The involvement of red teams indicates that the system is being stress-tested for safety weaknesses, which typically precedes the release of a new Claude version. This aligns with Dario Amodei’s public statements about Anthropic’s goal to refresh their AI models roughly every three months. The timing matches this cadence: Claude 4 launched a quarter ago, and last summer Sonnet 3.5 arrived around the same period, supporting the idea of an imminent update.

There’s speculation, reportedly from within Anthropic, that the next Claude release will carry the version 4.1 or 4.2 label rather than the rumored 4.5. This reflects Anthropic’s move to clarify versioning for users and developers, an area that’s become a focus as the model lineup expands. If past release patterns hold, the model could appear within three to four weeks—red teaming often lasts two weeks, and last time only a week separated the end of red team testing from public launch.

For enterprise customers and developers using Claude, this upgrade would likely be important for compliance-heavy and high-stakes use cases, where safety improvements and predictable updates matter. While details about Niptune v3’s changes are not public, the overall process points to Anthropic’s continued prioritization of robust safety layers as part of their commercial model rollout. The company has positioned itself as especially focused on safety in contrast to some competitors, which makes these kinds of staged, classifier-based systems central to its current strategy.