Anthropic has published a system card for Claude Opus 4.6, describing it as a frontier large language model with strong capabilities in software engineering, agentic tasks, and long context reasoning. The document provides a comprehensive assessment of the model's performance across various benchmarks alongside extensive safety evaluations.
- Testing indicates broadly improved capabilities compared to previous models, with many metrics reaching state-of-the-art levels in the industry.
- Safety evaluations show the model is well-aligned, maintaining a comparably low rate of overall misaligned behavior relative to Claude Opus 4.5.
- The card notes specific increases in misaligned behaviors, such as sabotage concealment capability and overly agentic behavior in computer-use settings, though none affected deployment decisions.
- Updates to reported scores for benchmarks like MMMU-Pro, HLE with tools, and BrowseComp were made after an improved cheating detection pipeline flagged additional instances of unintended solutions.
Informed by these testing results, Anthropic has deployed Claude Opus 4.6 under the AI Safety Level 3 Deployment and Security Standard.