Anthropic Releases Claude Opus 5.5 with 40% Cost Reduction and Record Alignment Scores

Anthropic has released Claude Opus 5.5. This model reduces execution costs for general workloads by 40% compared to its predecessor, Opus 5, under default settings. Input pricing is $4/M and output is $20/M, which is 20% lower than Opus 5, while cache read costs have been reduced by 60% to $0.20/M. Output generation speed has improved by more than 30%.

Performance and Benchmarks

Opus 5.5 ranked first in agentic coding, computer use, and knowledge work benchmarks. Key metrics include Terminal-Bench 4.0 at 66.4%, FrontierCode at 54.4%, and GDPval-AA Elo of 1846. Notably, in Medium effort mode, it demonstrated superior performance at approximately 1/5 the cost of GPT-6 Astra's max effort mode. In real-world tests, it matched the performance of Opus 5's high effort mode while using 20–25% fewer output tokens.

Alignment and Safety Evaluation

The model achieved the highest alignment scores to date in external evaluations and automated behavioral audits. In an audit of approximately 2,000 scenarios, it showed the lowest misalignment scores among recent Claude models, with attempts to bypass new container boundary tests reduced by approximately 85% compared to Opus 5. Biological and cybersecurity capabilities were assessed as similar to Claude Mythos 5.1, resulting in the application of the same Life Sciences/Cyber Verification Program safeguards as Fable 5.1. Prompt injection resistance is higher than Opus 5, and it recorded the lowest injection success rates alongside Fable 5.1 in the Gray Swan benchmark.

Deployment

The model is deployed across all platforms, including AWS, Google Cloud, and Microsoft Azure, and is accessible via the model ID claude-opus-5-5 on the Claude Platform. Verified organizations can use the model for biological research through the Life Sciences Verification Program, and cybersecurity practitioners can access it through the Cyber Verification Program.