Anthropic's most capable model, codenamed "Model 2," is for internal use only
Anthropic is running an unreleased internal model codenamed "Model 2" that outperforms every publicly available version of Claude The model is classified in the "Mythos" tier and scores approximately 1.5 points above Claude Mythos 5 on Anthropic's internal AECI capability index Despite being stronger overall, Model 2 is weaker in certain areas and does not represent a dramatic capability leap comparable to the Opus 4.6 to Mythos transition The model is used extensively for coding, data generatio
Analysis
TL;DR
- Anthropic is running an unreleased internal model codenamed "Model 2" that outperforms every publicly available version of Claude
- The model is classified in the "Mythos" tier and scores approximately 1.5 points above Claude Mythos 5 on Anthropic's internal AECI capability index
- Despite being stronger overall, Model 2 is weaker in certain areas and does not represent a dramatic capability leap comparable to the Opus 4.6 to Mythos transition
- The model is used extensively for coding, data generation, and research/engineering, often through continuously running agents
- Anthropic rates the overall misalignment risk as "low" and has no plans to release Model 2 externally at this time
Why It Matters
Anthropic's internal deployment of a more capable unreleased model signals the accelerating pace of capability gains happening behind closed doors, even as public releases remain incremental. The fact that Claude already writes most of Anthropic's production code underscores how deeply AI integration has penetrated core engineering workflows, making internal model performance a direct competitive differentiator. For the broader industry, this highlights the growing gap between what companies can do internally and what they choose to release publicly.
Technical Details
- Model Classification: "Model 2" is placed in the Mythos class, positioned slightly above Claude Mythos 5 on Anthropic's internal AECI (Anthropic Evaluation Capability Index), approximately 1.5 points higher
- Capability Profile: The model shows modest overall improvement rather than a dramatic jump; the gain is notably smaller than the transition from Mythos Preview to Mythos 5, and it exhibits weaknesses in certain unspecified areas
- Internal Use Cases: Heavily utilized for software coding, synthetic data generation, and research and engineering tasks, with deployment sometimes occurring through autonomous agents running continuously
- Safety Review: Underwent internal review prior to deployment but received less rigorous testing than Mythos 5; no new or more concerning misalignments were identified, and overall misalignment risk was rated "low"
- Release Status: No external release planned; the model remains strictly internal as of the August 2026 Risk Report
Industry Insight
- The incremental nature of Model 2's improvements—despite being Anthropic's most capable model—suggests that the easiest gains may have already been captured, and future progress could require fundamentally different approaches rather than scaling alone
- Anthropic's reliance on Claude for production code and the deployment of Model 2 through persistent agents reflects a broader industry trend where AI is moving from assistive tool to autonomous infrastructure, raising both productivity and safety questions
- The deliberate decision to keep a clearly superior model internal signals that capability withholding is becoming a strategic norm among top AI labs, which could slow public benchmark progress while widening the gap between leading labs and everyone else
Disclaimer: The above content is generated by AI and is for reference only.