Anthropic's risk report discloses an unreleased Model 2 more capable than Mythos 5

Anthropic has a model that is better than the one customers are allowed to use, and it says in writing that it has no plans to hand it over.

What the risk report says

The disclosure sits in the company’s August 2026 Risk Report, published on August 14 with a coverage date of July 15. The report introduces an unreleased internal system it calls Model 2 and describes it as “somewhat more capable than Mythos 5,” a “noticeable improvement on Mythos 5 for many tasks relevant to internal use,” though not a jump on the scale of the step from Claude Opus 4.6 to Mythos Preview.

Then the line that matters: “We do not currently have plans to release this model externally, and have not run all of our typical suite of predeployment assessments, so we have somewhat lower confidence in our beliefs about its capabilities.”

Model 2 and Mythos 5 are, in the report’s words, “used heavily within Anthropic for coding, data generation, and other agentic use cases.” The same document notes that Claude “authors a large majority of the code merged into our production codebases.” Separately, the report raised the company’s own misalignment risk assessment from very low to low.

Procedural, not alarming

The stated reason for holding Model 2 back is procedural rather than alarming. The company has not finished its usual predeployment evaluation suite, and it is explicit that this lowers its confidence in its own capability estimates. Nothing in the report describes a frightening result; it describes an unfinished checklist.

Still, the shape of the situation deserves attention. A frontier lab is running a model internally that is stronger than anything it sells, using it to write the code that builds the next one, and publishing that fact in a safety document rather than a product announcement.

The gap between the frontier you can rent and the one that exists

For most of the past few years, the distance between the best deployed model and the best internal model has been a matter of inference, guessed at from benchmarks. This report documents the gap directly, in the lab’s own words, with a named internal system and a stated reason for not shipping it.

That is useful information for anyone planning around model capability, and it carries a quiet implication: the public model is a product decision, not a capability ceiling. The questions the report does not answer are how long Model 2 stays internal, whether the full evaluation suite is eventually run, and whether disclosing unreleased systems becomes standard practice across labs or stays an Anthropic particularity.