Prompting Claude Opus 5.5

Claude Opus 5.5 processes output tokens over 30% faster and uses fewer tokens, reducing task completion costs and latency.

Anthropic has released documentation for Claude Opus 5.5, detailing significant advancements in its capabilities and prompting strategies. The new model processes output tokens over 30% faster than Claude Opus 5 and typically uses fewer tokens to complete tasks. Key improvements include enhanced agentic coding, where Opus 5.5 matches or surpasses Opus 5 at high effort with fewer steps and tokens, and sustains long-running autonomous work better. Knowledge work performance is also boosted, with fewer factual errors and better handling of complex documents and financial modeling. The model excels at interpreting visual inputs like charts and screenshots with greater accuracy and detail. A crucial change is that thinking is always on in Opus 5.5; prompts previously designed for disabled thinking in Opus 5 require adjustments. Effort calibration is now the primary control for balancing intelligence, latency, and cost, with Opus 5.5's medium effort often outperforming Opus 5's high effort. The API has four breaking changes from Opus 5, and prompts requiring the model to explicitly state its reasoning in the response may be declined.

The introduction of Claude Opus 5.5 signifies a substantial leap in LLM efficiency and utility, particularly for agentic tasks. Its improved performance in areas like autonomous code auditing and multistep workflows suggests a market shift towards more robust, less supervised AI operations. Developers can anticipate more reliable and cost-effective automated systems, capable of handling complex, long-duration projects with greater autonomy, potentially lowering operational overhead for businesses relying on AI agents.

The always-on thinking mechanism in Claude Opus 5.5, coupled with the ability to calibrate effort for performance tuning, indicates a move towards more granular control over LLM behavior. This allows for a finer balance between computational cost and output quality, crucial for real-time applications and complex problem-solving. Businesses integrating Opus 5.5 should re-evaluate their prompting strategies to leverage these new controls, ensuring optimal resource allocation and minimizing potential issues arising from previous prompting techniques.