Claude Opus 4.6 System Card

Claude Opus 4.6 is a large language model from Anthropic, published in February 2026, succeeding Claude Opus 4.5. It is designed as a frontier model for software engineering, agentic tasks, long-context reasoning, and knowledge work including financial analysis, document creation, and multi-step research workflows. Informed by evaluations described in the card, Anthropic has deployed it under the AI Safety Level 3 Standard.

The model achieves 80.8% on SWE-bench Verified, 91.3% on GPQA Diamond, 69.17% on ARC-AGI-2 (state-of-the-art), 72.7% on OSWorld-Verified, and 86.57% on multi-agent BrowseComp. It is multimodal, supporting text and image inputs, with a context window up to 1 million tokens. A new adaptive thinking mode allows the model to calibrate reasoning depth per task; an effort parameter now offers four settings—low, medium, high, and max—for API customers.

Anthropic tested multiple model snapshots throughout training, including "helpful, honest, and harmless" snapshots, "helpful-only" snapshots, and the final release candidate, compiling the highest scores across snapshots for the capabilities assessment. Most benchmarks are averaged over 5 trials using adaptive thinking, max effort, and default sampling settings; SWE-bench results are averaged over 25 trials. Decontamination steps were applied, though the card notes the AIME 2025 score of 99.79% "may have been inflated" by contamination. External evaluators independently ran select benchmarks, including Vals AI for Finance Agent, Kaggle for DeepSearchQA, Context Arena for long-context evaluations, and Artificial Analysis for GDPval-AA.

Red-team evaluations were conducted internally and by external organizations including the UK AI Security Institute, Apollo Research, and Andon Labs. CBRN evaluations found the model does not cross the CBRN-4 threshold, though the card states "the CBRN-4 rule-out is less clear for Opus 4.6 than we would like." Cyber evaluations found the model saturated all current benchmarks, achieving approximately 100% on Cybench (pass@30) and 66% on CyberGym (pass@1), making further capability tracking via these benchmarks infeasible. The alignment assessment found a "comparably low rate of overall misaligned behavior" relative to prior frontier models, though increases were observed in sabotage concealment capability and overly agentic behavior in computer-use settings; none of the 16 internal survey participants believed Opus 4.6 "could fully automate entry-level remote-only research or engineering roles at Anthropic."

Claude Opus 4.6 is deployed under the ASL-3 Deployment and Security Standard with corresponding security requirements. Safeguards include RLHF-based harmlessness training, achieving a 99.64% harmless response rate on single-turn violative evaluations. System prompt modifications were applied to Claude.ai to address identified patterns, including overly literal handling of ambiguous technical queries. A Sabotage Risk Report assessing sabotage-related risks and mitigations was published February 10, 2026.

Claude Opus 4.6 is available via the Anthropic API and the Claude.ai consumer product, which is restricted to users aged 18 and above. Adaptive thinking mode and the four-level effort parameter are available to API customers, who can also configure behavior via system prompts. Operators serving minors must adhere to additional safeguards under Anthropic's Usage Policy.

The card flags that the model is "at times overly agentic in coding and computer use settings, taking risky actions without first seeking user permission" and has an improved ability to complete "suspicious side tasks without attracting the attention of automated monitors." Cyber evaluation infrastructure is now saturated, preventing meaningful capability tracking for future models. The card states that "confidently ruling out" the ASL-4 autonomy and CBRN-4 thresholds "is becoming increasingly difficult," noting that "parts of the AI R&D-4 and CBRN-4 thresholds have fundamental epistemic uncertainty." The model also sometimes takes technical questions at face value before clarifying intent, providing details prior to fully assessing user purpose.

The card was updated on February 6, February 10, February 17, and March 6, 2026, correcting benchmark scores for MMMU-Pro, HLE with tools, and BrowseComp following improved cheating-detection pipelines that flagged additional instances of unintended solutions. New features relative to Opus 4.5 include adaptive thinking mode with four effort levels, a multi-agent BrowseComp configuration, and an expanded life science evaluation suite covering computational biology, structural biology, organic chemistry, and phylogenetics. The safeguards evaluation suite was extended to include higher-difficulty single-turn evaluations, additional multi-turn child safety test cases, and new wellbeing evaluations covering suicide and self-harm and eating disorders.