Claude Opus 5.5 generates output tokens over 30% faster than Claude Opus 5 and often finishes tasks using fewer tokens. Existing prompts remain effective, but specific patterns address behavioral shifts. Calibrate effort when runs cost more or take longer. Adjust prompts if thinking was previously disabled. For unattended agents that stop mid-task, review progress reporting. Handle refusals by refining safeguard triggers. Implement user-facing progress updates for long agentic turns. In multi-app workflows, explicitly guide context exploration. Use time signals to optimize multiagent harnesses. Modify chat system prompts to manage lengthy initial thinking. Clearly mark pasted text within user messages to prevent instruction confusion. Employ tools for analyzing dense charts, diagrams, or screenshots to improve accuracy. Apply frontend design defaults to avoid generic output.

Claude Opus 5.5 excels in agentic coding and code review, matching or exceeding Claude Opus 5 at high effort while using fewer steps and tokens. It sustains long-running autonomous work like multi-hour audits and migrations. It catches more bugs during code review with fewer false alarms and explains changes clearly. For knowledge work, it reduces incorrect figures and misattributed sources, performing better in financial modeling and detail detection within large inputs. Produced spreadsheets, slides, and documents require less editing. Communication is direct, clearly stating actions, findings, and requirements. Visual interpretation is significantly improved; even at lowest effort, it reads dense charts more accurately than Claude Opus 5 at highest effort, using fewer tokens. It also better understands positional relationships in diagrams.