Anthropic Discloses 'Model 2' Outperforming Claude Mythos 5 in Risk Report

Anthropic's August 2026 risk report has disclosed a new, unreleased AI model named "Model 2" that slightly outperforms their current strongest model, Claude Mythos 5, on internal benchmarks like "CoBench v2". Users speculate that Model 2 and its predecessor Model 1 are further iterations of the Claude Mythos Preview model. This disclosure provides a rare glimpse into Anthropic's internal model roadmap and safety evaluations, showing steady incremental progress beyond their highly restricted frontier models. It highlights how AI developers are managing risk assessments for next-generation models before public deployment. While Model 2 shows superior performance across multiple tasks on the "CoBench v2" benchmark compared to Claude Mythos 5, the performance gains are relatively modest and focus on improving general capabilities.