TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Anthropic published guidance on adapting prompts and integrations for Claude Opus 5.5, covering effort calibration, always-on thinking, agent workflows and visual tasks. The company says the model generates output tokens more than 30% faster than Claude Opus 5, but advises testing settings against each user’s own evaluations.
Anthropic has published prompting guidance for Claude Opus 5.5, describing how developers can adjust effort settings, API integrations and agent instructions for the model. The guide says Opus 5.5 generates output tokens more than 30% faster than Opus 5 and often completes the same task with fewer tokens, while warning that settings carried over from the previous model can affect latency, cost and response length.
The documentation recommends starting with medium effort, Opus 5.5’s default, and testing several effort levels against a team’s own evaluations. Anthropic says effort names do not represent the same amount of thinking across models: in its testing, Opus 5.5 at medium matched or exceeded Opus 5 at high on coding and knowledge-work evaluations. The company also reports that low effort came close on several coding evaluations at lower cost. These are Anthropic’s test results, not an independent comparison.
The guide says thinking is always on in Opus 5.5, making effort the main control for how much the model thinks. Developers who previously used high effort or disabled thinking for Opus 5 may see longer turns and more output tokens after switching. Anthropic advises setting the token limit high enough for both thinking and the final reply, since thinking tokens count toward max_tokens even when the integration does not return the thinking content. It says a 128,000-token limit worked well in its testing for long agentic coding tasks, and recommends reserving xhigh and max effort for work where evaluations show a quality gain.
Other sections address unattended and multiagent tasks, progress updates, refusals, chat system prompts, pasted text, multi-app workflows and visual inputs such as charts and screenshots. The guide also includes advice for frontend design and suggests using tools for complex visual inputs. Anthropic says existing Opus 5 prompts should generally work and that its Opus 5 prompting guide remains a reasonable starting point, with targeted changes based on observed behavior.
Settings Shape Opus 5.5 Workflows
The guidance matters to teams upgrading applications or agents because a model change can alter latency, token use and task completion even when prompts stay the same. Always-on thinking means token limits sized for a prior setup with thinking disabled may cut off a reply. Anthropic’s advice to run evaluations at multiple effort levels gives developers a way to compare quality and cost in their own workloads rather than assuming the same setting will transfer.
The document also focuses on how Opus 5.5 behaves during longer tasks: whether an unattended agent stops early, whether users receive progress reports, and whether agents gather relevant information across connected apps. Those details can affect the reliability of deployed systems as much as the wording of a single prompt. Anthropic’s performance claims may help explain why it recommends revisiting settings, but the guide does not establish that every workload will see the same gains.
How the Guide Fits Opus 5.5
Anthropic presents this page as a prompting guide specific to Opus 5.5, alongside separate documentation for the model’s capabilities and API changes and advice that applies across current Claude models. It says prompts written for Opus 5 should remain a reasonable starting point, so the new guide is chiefly a set of adjustments for cases where users observe different behavior.
The guidance highlights differences in default effort: Opus 5.5 defaults to medium, while Opus 5 defaults to high. Anthropic says Opus 5.5 may think more per turn at a given effort level, especially at xhigh and max. It therefore cautions against simply carrying a previous effort setting forward without measuring its effect.
Questions the Guide Leaves Open
The supplied documentation does not identify a publication date, provide independent verification of Anthropic’s performance comparisons, or give results for every task type and deployment. It is unclear how broadly the reported speed and token savings apply to individual applications, since workloads, prompts and system settings differ.
The guide points developers toward their own evaluations but the source material does not provide a universal effort setting for each use case. Nor does it establish that Opus 5.5 will always outperform Opus 5 at lower effort. Actual cost and response times will depend on the model’s thinking, the task, token limits and the integration.
Test Settings Against Your Workload
Anthropic’s guidance points developers toward testing effort levels against their own evaluations and adjusting token limits to account for thinking as well as replies. Teams using Opus 5 prompts can start from those prompts, then consult the relevant section of the guide when they observe longer turns, early stops, refusals or missed visual details.
The documentation does not announce a scheduled follow-up or a future release milestone. For now, the next practical step for developers is to measure how Opus 5.5 performs in their applications and update prompts or harness settings where those results call for changes.
Key Questions
What did Anthropic publish?
Anthropic published a Claude Opus 5.5 prompting guide covering effort settings, thinking behavior, agent workflows, visual inputs and other integration patterns.
What effort setting does the guide recommend?
It recommends starting at medium effort, Opus 5.5’s default, and comparing settings using evaluations built around the team’s own tasks.
Why might an Opus 5 integration need a higher token limit?
Anthropic says thinking is always on for Opus 5.5 and that thinking tokens count toward max_tokens, even when the integration does not return the thinking content. A limit set for an earlier configuration may leave too little room for the final reply.
Do existing Claude Opus 5 prompts still work?
Anthropic says existing Opus 5 prompts should perform well without changes and remain a reasonable starting point. It recommends targeted adjustments when users observe differences in their own applications.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
