Claude Opus 5

Anthropic's model for complex agentic coding and long-horizon work. These notes describe model behavior, not Claude Code behavior.

• Give difficult multi-file or end-to-end work a complete specification up front, including scope, constraints, and the intended stopping point.
• Use effort deliberately: start with the default and evaluate lower effort for cost- or latency-sensitive work; use xhigh for demanding coding or agentic tasks when the calling surface supports it.
• Prompt for the response length, tone, and written-deliverable density you need; lowering effort changes thinking more reliably than visible response length.
• In code review, state whether the first pass should maximize finding coverage or filter to a severity bar; Opus 5 may follow a narrow reporting instruction literally.
• Keep validation instructions proportional to the task. The model already self-checks, so add targeted checks when they protect a concrete requirement.

• Leaving a narrow task's scope implicit when extra analysis or cleanup would be costly.
• Repeating generic "double-check" or mandatory subagent-verification instructions without a specific risk to cover.
• Assuming model-level self-correction replaces the tests, tools, or approvals supplied by the harness.

Claude Opus 5 is documented with a 1M-token context window and adaptive thinking on by default. Tool access, subagent orchestration, and approval behavior still come from the harness and its configuration.

When thinking is disabled, Anthropic documents additional constraints: disabling thinking at xhigh or max returns an error, and visible output can occasionally contain a tool call or internal XML tags. Keep thinking enabled where practical and validate API-specific behavior in the integration being used.