Anthropic Releases Prompting Guide for Claude Opus 5.5
▼ Summary
– Anthropic released a prompting guide for Claude Opus 5.5, advising developers to retest effort levels rather than reusing settings from the previous Opus 5 version.
– The new model defaults to medium effort instead of high, and removes the ability to disable thinking, which previously caused errors when attempts were made to turn it off at lower efforts.
– Chat applications are encouraged to remove system prompts instructing the model to think carefully, as the model now manages its own reasoning depth based on the selected effort level.
– Agent teams should utilize time budgets to improve research task efficiency, with tests showing that small groups using time signals outperform solo agents without such constraints.
– Users inputting external text should employ tags with random IDs and specific system notes to mitigate prompt injection risks, acknowledging this provides only limited protection.
Anthropic has released updated guidance for developers integrating Claude Opus 5.5, urging a shift in how models are prompted to balance performance, speed, and cost. The new model, which launched on September 22, defaults to a medium effort setting rather than the high effort used by its predecessor, Claude Opus 5. This change requires teams to re-evaluate existing configurations, particularly those that previously relied on explicit instructions for the model to “think carefully” before generating responses.
Adjusting Effort Levels and Defaults
The most significant structural change in Opus 5.5 is the default behavior regarding computational effort. While Opus 5 operated at high effort by default, Opus 5.5 runs at medium effort out of the box. This adjustment allows applications to achieve faster response times without requiring immediate code modifications, though it does necessitate a review of current settings. Anthropic’s documentation emphasizes that effort is the primary lever for tuning the trade-off between quality, latency, and expense.
Testing indicates that Opus 5.5 operating at medium effort performs on par with or better than Opus 5 at high effort for tasks involving coding and knowledge work. Consequently, the guide advises developers to lower effort settings as a first step to reduce thinking overhead, reserving higher tiers such as xhigh and max for complex scenarios where superior quality justifies the additional time and cost. Notably, the ability to disable thinking entirely, which was available in Opus 5 at high effort or below, is no longer an option. Requests attempting to turn off thinking will return an error, reinforcing the model’s internal decision-making process regarding depth of analysis.
Rethinking System Prompts for Chat Applications
A key recommendation in the new guide concerns system prompts that instruct the model to deliberate extensively. Anthropic found that removing lines commanding the model to “think carefully” accelerated response generation in chat products without causing a measurable drop in output quality. The statement remains: “no clear decline in the quality of the reply.”
This shift reflects Opus 5.5’s autonomous approach to reasoning. The model now determines the appropriate depth of thought based on the assigned effort level, rendering manual instructions for careful consideration largely redundant. While previous documentation for models like Claude Opus 4.7 suggested adding such directives when keeping effort low to maintain speed, the newer architecture handles this internally. Developers are encouraged to strip these phrases from their chat application prompts to streamline interactions.
Optimizing Agent Workflows and Input Handling
For multi-agent systems, the guide introduces strategies for managing time budgets. Anthropic’s tests demonstrated that small groups of agents utilizing elapsed time signals completed research tasks more efficiently than single agents working without such constraints. These groups maintained answer quality comparable to solo performers while operating under advisory time limits. The documentation notes that while strict timeouts can be helpful, they may slightly reduce thoroughness; therefore, the time budget should be viewed as a suggestion rather than a hard stop.
Additionally, the guide addresses security concerns related to pasted text from external sources, such as emails. To mitigate risks like prompt injection, Anthropic recommends using tags with random IDs to mark incoming content and providing specific system instructions on how to handle this tagged text. This method serves as a protective layer, though the company clarifies that because tags are plain text, they offer only partial defense against sophisticated attacks.
Future-Proofing Integrations
Developers must also account for changes in output handling. Using an output cap (max_tokens) configured for Opus 5 with thinking disabled may result in truncated responses in Opus 5.5, as the hidden thinking process consumes part of the token limit. Finally, for frontend integrations, the guide suggests defining clear stylistic parameters to avoid generic AI aesthetics, noting that vague requests often lead to unintended default styles rather than customized designs. Existing Opus 5 prompts remain functional, but these specific adjustments ensure optimal performance with the latest iteration.
(Source: Search Engine Journal)
