Anthropic discloses internal, more-capable 'Model 2' it has no plans to release

Anthropic disclosed an internal, more-capable AI model it calls “Model 2” that it has no plans to release, in a Risk Report the company published on its own site on August 14, 2026. Model 2 is somewhat more capable than Mythos 5, currently Anthropic’s top-of-line public model.

The disclosure is unusual: an AI developer describing a system stronger than anything it sells, while nudging its own risk assessment upward. “We do not currently have plans to release this model externally,” the report said. Both Mythos 5 and Model 2 are used heavily inside Anthropic for coding, agentic work and data generation.

In the same August 2026 Risk Report, Anthropic raised its estimate of misalignment risk from “very low” to “low” and said Model 2 showed a noticeable improvement on many internal tasks, with early signs of acceleration in automated AI research-and-development capabilities.

One widely circulated figure could not be verified. Unattributed social commentary put Model 2 at 12.5 percentage points above Mythos 5 on an internal benchmark called CoBench v2. Anthropic’s report does not publish that number, and there is no independent way to confirm it; internal benchmarks are defined and scored by the vendor, not an outside evaluator.

The disclosure lands as Anthropic prepares to go public. The company’s preliminary second-quarter revenue reportedly topped $11.5 billion ahead of an expected October listing, a valuation backdrop that raises the stakes for how candidly it discusses capability and risk. A more-capable model kept in-house is a governance choice as much as a technical one, and the report frames it that way.