Anthropic built a model more capable than Mythos, but isn’t ready to release it

Anthropic has no plans to release an internal model known as “Model 2,” which appears to outperform its flagship Mythos model, even as the company continues pushing ahead with AI development.

Axios reports, citing Anthropic’s latest risk assessment, that the company routinely trains and evaluates multiple experimental models that are never intended for public release, with Model 2 being one of them.

The report also shows that Anthropic still considers the risk of severe harm from its models to be low, although its assessment has increased since the previous review. The company raised its overall estimate of misalignment risk in high-stakes situations from “very low” to “low,” pointing in part to recent cybersecurity incidents. Its models are also showing signs of improving their ability to conduct research and development autonomously, which could accelerate technical progress while increasing the potential for misuse.

According to the report, Model 2 shows a “noticeable improvement” across many of Anthropic’s internal tasks. The company is already actively using it alongside Mythos 5 for coding, agentic work and data generation, although the improvement is not as dramatic as the jump from Opus 4.6 to Mythos earlier this year.

“We do not currently have plans to release this model externally,” Anthropic says in the report.

Model 2 is an internal experimental system and is separate from the publicly available versions of Claude.

Anthropic also says it is less confident in its assessment of Model 2 than it was when evaluating models for previous reports, as some of its most concrete task-based tests are becoming less effective at capturing improvements in model capabilities.