Anthropic’s Secret Model is Even Better Than Mythos 5 But You Won’t Get It

Anthropic has disclosed the existence of an unreleased AI system, dubbed Model 2, which the company claims to be marginally superior to Claude Mythos 5. Model 2 was referenced in Anthropic's August 2026 Risk Report, though it remains unpublished. As of the July 15 coverage, Model 2 was among three frontier or near-frontier models yet to be made public.

When compared to Mythos 5 on Anthropic's internal CoBench v2 benchmark, which assesses 449 real-world research and engineering challenges, Model 2 achieved 62.8%, outperforming Mythos 5 (50.3%) and Mythos Preview (54.8%). In contrast, Claude Opus 4.7 scored 27.4% and Opus 4.6 attained 15.6%. Anthropic estimates that a model capable of fully replacing its technical research team would require a CoBench score of at least 85%, indicating that Model 2 is still far from this mark.

The company describes Model 2 as a "noticeable improvement" over Mythos 5 for numerous internal tasks, such as coding, data generation, research, and engineering, including persistent AI-agent deployments. However, Anthropic has no intention to release Model 2 to the public. They have not yet completed their standard predeployment assessments, leading to lower confidence in their assessment of the model's capabilities compared to fully evaluated models.

Additionally, Anthropic's internal deployment review found no new or more concerning forms of misalignment compared to those observed in Mythos 5. There is currently no evidence suggesting that Model 2 will eventually launch as Claude Mythos 6 or under any other commercial name.