Anthropic has published the system card for Claude Opus 4.5, a frontier large language model designed with prominent capabilities in software engineering, tool use, and computer use. The document provides a comprehensive assessment of the model's technical abilities alongside extensive safety evaluations covering safeguards, honesty, agentic safety, and alignment.
- Testing identified several state-of-the-art capabilities across benchmarks including SWE-bench, Terminal-Bench, ARC-AGI, and WebArena.
- The model was assessed for broad alignment, showing low rates of undesirable behavior such as sycophancy, deception by omission, and sabotage propensity.
- Safety evaluations included tests for malicious use of agents, prompt injection risks, and compliance with the Responsible Scaling Policy regarding CBRN (chemical, biological, radiological, and nuclear) risks.
- The system card clarifies training data usage, noting that while ARC-AGI-1 scores were updated to reflect semi-private test sets, the model was not trained on the public train/test split of that dataset.
The report concludes that Claude Opus 4.5 is a broadly well-aligned model and has been deployed under the AI Safety Level 3 Standard.