Anthropic positions Claude Opus 5.5 as its flagship for complex coding, research, and long-running professional work. The company says it performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5 on typical tasks. Its token prices are 20% lower than the previous model, and it often uses fewer tokens to finish work. The announcement includes early examples in code migration, website optimization, and extended office tasks. Read Anthropic’s release notes.

Official image for Claude Opus 5.5 发布，输入输出单价下调 20%
Official image from Anthropic; click to open the original source.

Where the savings come from

Standard API rates are $4 per million input tokens and $20 per million output tokens, 20% below Opus 5. Cached reads cost $0.20 per million tokens. Anthropic separately estimates a 40% reduction in total cost on typical workloads with default settings. These are different measures: some savings come from lower token rates and some from using fewer tokens to complete a task. Your actual result depends on codebase and document size, tool calls, and retries.

Fast mode has separate pricing: $8 per million input tokens and $40 per million output tokens. It may be useful when response time matters, but a background batch job should be tested to see whether the added cost is justified. Published benchmarks cannot guarantee the same savings for your team.

Evaluate difficult work with a controlled test

  1. Choose reproducible tasks such as a cross-module code migration, a performance regression, or evidence synthesis across long documents.
  2. Keep tools, context, and acceptance criteria fixed, and run Opus 5 and 5.5 on the same tasks.
  3. Record correctness, test pass rate, human review and rework time, and token cost in standard and Fast modes.
  4. Let the model propose and test changes in a temporary branch or sandbox. Keep engineers in charge of merge, release, and consequential edits.

Safety and availability

Anthropic reports external evaluations and its automated behavioral audit, saying Opus 5.5 is better at respecting boundaries and resisting prompt injection, while noting that it can still make mistakes. Because of its capabilities in biology and cybersecurity, some research uses require additional verification; eligible organizations can apply to the Life Sciences Verification Program, with Cyber Verification access expanding separately. These restricted high-risk permissions are distinct from ordinary development and knowledge work. Check regional availability, model names, and limits separately for Claude, the API, and cloud platforms.