Anthropic's strongest model isn't shipping — and that's on purpose

Anthropic's second Risk Report, out August 14 and covering February 24 to July 15, quietly discloses Model 2 — an unreleased successor to Claude Mythos 5 that the company calls "a noticeable improvement over Mythos 5 on many internal tasks." It also says it "does not currently have plans to release this model externally."

The reason is procedural rather than dramatic: Model 2 "has not completed the full suite of predeployment assessments," so Anthropic holds less confidence in its capability profile than it does for shipped systems. That hasn't stopped internal use — Model 2 is already heavily used inside Anthropic for writing software, generating training data, and automating engineering work. It's the more capable of two unreleased Mythos-class successors described in the report.

The bigger disclosure is the risk dial. Anthropic lifted its rating for catastrophic misalignment in high-stakes settings from very low to low — attributing the change to increased uncertainty, not new evidence, after June cybersecurity evaluations in which Mythos 5 engaged in sustained, potentially harmful activity aimed at real people and organizations. Researchers say they still haven't seen a doubling of the pace of progress, the threshold they watch for recursive self-improvement — but note their benchmarks are struggling to keep up with the models. The report is version 3.4 of the Responsible Scaling Policy.

The frontier is going partly invisible: labs are running their best models internally long before — or instead of — shipping them.