The latest version of Anthropic’s conversational AI assistant, Claude Opus 5.5, now comes with a new guide. It is designed for developers looking to integrate the model into their own systems, and it covers a wide range of topics, including effort calibration and multi-agent workflows.
This guide is titled “Prompting Claude Opus 5.5,” and it centers on comparing how Claude Opus 5 behaves against the new release. It presumes readers already know the older version’s workings, so it skips over the fundamentals and moves directly to the updates instead.
Calibrating Effort
The key difference lies in how much effort is put in. Claude Opus 5.5 produces tokens at a rate over 30 percent faster than its predecessor, and it usually completes the same task using fewer tokens. As a result, current prompts will generally continue to function without needing to be altered, even though the underlying thought process has shifted.
The recommendation from Anthropic is to begin with medium effort, which serves as the default on Claude Opus 5.5. The earlier edition instead set its default at high effort, making the two settings not interchangeable. The guidance suggests running tests across several levels alongside your own evaluations rather than simply keeping the setting you used on Claude Opus 5.
The guide notes that effort level names do not correspond to the same amount of thinking across models. In Anthropic’s testing, Claude Opus 5.5 at medium effort matches or exceeds Claude Opus 5 at high effort on coding and knowledge-work evaluations, and low effort comes close to it at much lower cost.
Thinking Behavior
Claude Opus 5.5 thinks more per turn than its predecessor, especially at xhigh and max. That means a prompt that worked fine at a given effort level on the older model may now produce longer turns and more output tokens. The guide offers three fixes:
- Set max_tokens high enough to leave room for the model’s thinking tokens and the reply.
- Reserve xhigh and max for work where you’ve measured a quality gain.
- Lower effort first if you want less thinking.
Thinking counts toward max_tokens even when thinking content isn’t returned to you, so a limit sized for Claude Opus 5 with thinking off can cut replies off. For long turns that agentic coding can produce, a max_tokens of 128,000 has worked well in Anthropic’s testing.
Safeguard Refusals and Unattended Runs
The guide describes a stop_reason of “refusal” as falling under safeguard refusals. When requests come back with this reason, it signals a long agentic turn that looks silent. You may find yourself wanting updates at predictable points during these extended periods.
Among the issues covered by the guide are unattended agents that cease work midway through a lengthy assignment, having first reported on their progress. Such instances are referred to as unattended agentic runs. In each case, the agent stops short of completion while still in the middle of its assigned task.
Multi-Agent Harnesses
The guide offers time signals as a way to get a team of agents finished sooner. It involves coordinating multiple agents working on a shared task, each contributing a piece.
According to the guide, early testers reported stronger code review, catching more bugs than on Claude Opus 5, while also generating fewer false alarms. The model lays out its changes in clear, easy-to-understand terms.
Visual Inputs
Anthropic’s testing found that the model reads dense charts more accurately than Claude Opus 5, even at its lowest effort setting. This happened while using a small fraction of the output tokens compared to how Claude Opus 5 performs at its highest setting.
Meaning often depends more on arrangement than on words alone: which boxes an arrow joins in a flowchart, what has changed between two editions of a diagram, or precisely when a meeting begins and ends in a calendar snapshot.
Knowledge Work and Agentic Coding
The model excels at tasks like constructing a financial model and one-page summary for a transaction, or locating and correcting errors in a valuation workbook. A real code base is where the model performs best on multistep tasks. It guides a change all the way through a large code base until its tests pass, and it can sustain long-running autonomous work longer than Claude Opus 5.
The system handles long-running audits and code migrations across large code bases using parallel subagents, with minimal supervision. According to the guide, it matched or surpassed Claude Opus 5 on these demanding, high-effort jobs, completing them in fewer steps and consuming fewer tokens.
Past Text, Context, and Frontend Design
Inside the message a user pasted, there is text containing instructions the model follows. That text includes a particular style of marking pasted content, which the guide describes. The model can manage requests that reach across several linked apps, although it might miss details the original request failed to mention.
Frontend output looks generic, and the guide covers frontend design defaults.
The Key Facts
- Claude Opus 5.5 generates output tokens more than 30 percent faster than Claude Opus 5
- Claude Opus 5.5 tends to finish the same task with fewer tokens
- Medium effort is the default on Claude Opus 5.5
- Low effort on Claude Opus 5.5 matches high effort on Claude Opus 5 at much lower cost
- A max_tokens of 128,000 has worked well in Anthropic’s testing for long agentic coding turns
- Requests can return a stop_reason of “refusal”
- The model reads dense charts more accurately than Claude Opus 5, even at its lowest effort setting
This document describes how the two versions behave differently, explained with practical detail throughout its pages.
Source material: “Prompting Claude Opus 5.5,” claude.com.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.