Claude Opus 5.5

What is Claude Opus 5.5?

Claude Opus 5.5 is Anthropic’s model for demanding coding and knowledge work. OpenRouter lists text, image and file inputs with text output. Consider it for a difficult review that needs sustained reasoning across several pieces of evidence; the API model is separate from the Claude application and its subscription.

Reserve it for a difficult decision

Anthropic positions Opus 5.5 around long-running coding and knowledge tasks. A useful assignment is to review a cross-file change: provide the patch, surrounding code and intended behavior, then ask for concrete failures with the input that would trigger each one.

For a dense technical document, combine the text with the relevant diagrams and ask the model to identify contradictions or missing assumptions. Request source locations and distinguish an observed inconsistency from a hypothesis. This gives a reviewer something specific to verify.

Suggested workflow: review one risky change

Write down the behavior that must remain true. Supply a bounded patch and relevant tests, and ask the model to prioritize issues that change correctness or user-visible behavior. Require a reproducible example for each proposed defect.

Run or construct the suggested reproductions and discard unsupported findings. Fix the confirmed issue, rerun the relevant checks and review the final diff. Compare the time saved with the effort needed to inspect false alarms before expanding this approach to every change.

Pricing and always-on thinking

OpenRouter uses anthropic/claude-opus-5.5. At the October 17 check, the listed standard text rates were $4 per million input tokens and $20 per million output tokens. Caching, tools and processing options can change total charges; use current endpoint rates for a real estimate.

Anthropic documents always-on adaptive thinking for Opus 5.5, with effort controlling its depth, and says forced tool use (tool_choice set to any or to a named tool) is not supported and returns an error. A request configuration copied from an older model may therefore fail. Start with a small compatibility check before sending a large document or handing it a lengthy agent task.

Compare against a simpler first pass

For a straightforward summary or fixed-label classification, test GPT-6 Luna first. Use the same acceptance criteria on both candidates and compare actual failures, review time and cost. Opus is more interesting when the task contains a difficult reasoning step that a cheaper first pass did not resolve.