Anthropic‘s internal Risk Report from August 2026 confirms the company operates an unreleased model named Model 2 that outperforms every publicly available version of Claude. This internal tool sits in the Mythos class and scores approximately 1.5 points higher on the company’s AECI than Mythos 5, though it lacks the dramatic capability leap seen between Opus 4.6 and Mythos. While the performance gain is modest compared to previous major upgrades, the system handles significant internal tasks including coding, data generation, and continuous engineering research. Anthropic currently uses the model heavily through autonomous agents that run without constant human intervention. The company completed an internal review before deployment but did not subject the system to the same testing rigour as Mythos 5. Despite this lighter scrutiny, Anthropic rates the risk of misalignment as low and finds no new or more concerning safety issues. There are no plans to release Model 2 externally at this time.
The existence of this unpublished system matters because it reveals how the company manages its own development before public launch. Internal tools often set the standard for future public releases, meaning users may see capabilities that were already refined in private environments. Anthropic maintains a separate risk assessment for internal models, which allows faster iteration but creates a distinct testing pathway from public products. The company balances the need for rapid internal improvement against the requirement for thorough public verification. This approach suggests future public updates may benefit from earlier private refinement cycles.
- Model 2 scores 1.5 points higher than Mythos 5 on the AECI
- Internal testing was less rigorous than the public Mythos 5 release
- There are no current plans for an external launch
