🤖 The “Prompting Claude Opus 5.5” guide has been released
Claude Platform has published a set of prompting patterns for claude-opus-5-5. Adaptive thinking is always enabled, and the depth of reasoning is set by effort, defaulting to medium; on medium, the model does not lag behind Opus 5 on high in coding and knowledge benchmarks, generating tokens more than 30% faster.
🌍 Instead of manual budget_tokens and forced tool_choice in the API — a single effort knob. When migrating from Opus 5, requests with thinking:disabled will return a 400 invalid_request_error, and computer_20251124 has been replaced by computer_toolset_20260801.
👤 For agentic coding: effort: medium and max_tokens 128000 — thinking is counted in max_tokens. Choose response blocks by the type field, and reduce “reasoning” via effort, not through the prompt.
Source 1: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5-5 Source 2: https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5