Anthropic tells developers that a Claude Opus 5 prompt at high effort should usually run on Opus 5.5 at medium instead, and publishes a fresh prompting guide
Anthropic published a new prompting guide for Claude Opus 5.5 that tells developers to drop the effort setting from high to medium by default, remove instructions that told the older model to think before answering, and mark pasted user text with same-id tags so the model treats it as data rather than commands.
Image: Anthropic
Why it mattersA team that carries an Opus 5 prompt across without changes is paying for more thinking than the newer model needs and getting slower first tokens on top, so the fix is a config change and a few prompt lines rather than a rewrite.
Any developer who copied a Claude Opus 5 prompt straight into a Claude Opus 5.5 call is paying for more thinking than the newer model needs. Anthropic said so this morning in a prompting guide it published for Opus 5.5, urging teams to drop the effort setting from high to medium and to walk through eight prompt patterns before running the old prompt on the new model.
The guide starts from a single measurement, in Anthropic's own testing: Opus 5.5 at medium effort matches or beats Opus 5 at high on coding and knowledge-work evaluations, and on several coding benchmarks low comes close at "much lower cost". Because effort-level names do not carry across models, Anthropic recommends setting medium explicitly and only moving to xhigh or max when a quality gain has been measured.
Four prompt lines to change or add
The guide names three specific edits worth making before running an old prompt on the new model.
First, remove any "think carefully before answering" instruction from chat system prompts. Because Opus 5.5 always thinks and effort is the control, the instruction only pushes the model to think longer and makes replies start slower. Anthropic's chat-product test showed replies started sooner after the line came out, with no measurable quality decline.
Second, mark pasted text in user messages. To resist instructions that arrive through content a user copied from somewhere else, Anthropic tells developers to wrap each pasted block in an opening and closing tag carrying the same short random ID generated by the application, then add a system-prompt note telling the model to treat the text as data unless the user's own message asks for it.
Third, raise max_tokens. Thinking counts toward the token limit even when the thinking content is not returned, so a limit sized for Opus 5 with thinking disabled can cut a reply short on Opus 5.5. For long agentic coding turns, Anthropic says a max_tokens of 128,000, the model's ceiling, has worked well in its own testing.
The unattended-agent trap
Opus 5.5 writes short user-facing progress notes between tool calls, and some of those notes end the turn with text rather than a tool call, returning stop_reason: "end_turn". An unattended agent loop that reads a text-only turn as the end of the task will stop running there. The guide gives a system-prompt paragraph, written to be pasted into an unattended harness, that names four early-stop patterns to avoid: a long summary that closes by announcing the next step, an offer to continue unless the user prefers otherwise, a list of decisions for the user when none of them blocks the work, and stopping because the turn has been long or a milestone is done.
Two other paragraphs in the guide address the same category of long-agent problem: a system-prompt line that tells the model to explore several apps before acting when a workflow spans them, which Anthropic says lifted correct completion on multi-app automation tasks at both medium and max effort, and an "elapsed 340s / 1200s" time budget appended to each message that Anthropic reports made small research-agent teams finish sooner without a measurable quality decline.
The version of this guide already published for Claude Opus 5 remains "a reasonable starting point", so the shortest path for a team already running Opus 5 in production is to drop the effort setting, cut the "think carefully" line from chat prompts, add the pasted-text tags where users paste content, and, if agents run unattended, paste in the four-stop paragraph.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
