Skip to content
AI.info

The Pulse

Anthropic advises deleting "think hard" lines from Claude Opus 5.5 prompts

Anthropic says removing instructions to think hard made Claude Opus 5.5 start responding sooner in its tests, with no clear drop in quality. The company recommends revisiting effort settings and other habits developed for older models.

Anthropic advises deleting "think hard" lines from Claude Opus 5.5 prompts

Removing instructions to think hard made Claude Opus 5.5 start its replies sooner in tests, without a clear drop in quality, according to Anthropic. The company recommends revisiting prompting habits developed for older Claude models, including how much effort users ask the model to apply.

Opus 5.5 is the default model in Claude on Pro, Max, Ultra and Team plans. In a November 12 article, Anthropic recommended it for routine research, analysis and the production of documents, presentations and spreadsheets.

Effort and old prompts

Thinking is always on for Opus 5.5, with the amount determined by the request and the selected effort level. Anthropic recommends starting at Medium, the default, and moving to High when Medium stalls; higher settings can take longer and cost more without always improving the result.

In its coding and knowledge-work evaluations, the company says Opus 5.5 at Medium matched or beat Opus 5 at High. Its platform documentation also cautions that effort levels do not correspond to the same amount of thinking across models: at a given level, Opus 5.5 tends to think more per turn, particularly at the highest settings.

That does not mean all old prompts need rewriting. The platform guide says existing Opus 5 prompts should perform well without changes, while recommending that developers test effort levels rather than carrying over their previous setting.

In another test reported in the article, Factory technical staff member Zimu Li said Opus 5.5 at Medium matched Opus 5 at High while using 20 to 25% fewer output tokens.

Describe the finished work

Rather than giving instructions one at a time, Anthropic recommends describing the goal, desired output and points at which the model should stop and ask for input in the first message. It also suggests asking for a finished file rather than an outline when the intended result is a document or spreadsheet.

The company recommends attaching charts or diagrams directly instead of transcribing their values, a workaround it says may no longer be needed. For design requests, it says naming specific patterns to avoid, such as cream backgrounds or pill-shaped buttons, works better than a general request to avoid a generic look.

Connectors can give Claude access to work files and systems rather than relying only on material pasted into a prompt. On Team and Enterprise plans, an owner must enable a connector before members can use it. Anthropic also suggests supplying brand or writing guidelines, or pointing Claude to examples, while allowing for several rounds of refinement.

Prices and test results

For usage billed by the token, Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Sonnet 5.5 is priced at $2 and $10 respectively. Higher-priced models use plan limits faster on Pro, Max, Ultra and Team.

Anthropic says a more capable model can sometimes complete a task with less work, potentially offsetting much of its higher per-token price. It recommends Sonnet for quicker, well-scoped tasks with a clear brief.

In an internal research test, Opus 5.5 produced 18 reports on a company's quarterly performance using a saved web snapshot in which the earnings release was hard to locate. An automated grader checked figures and quotations against the snapshot, and 16 reports passed across effort settings; a report failed if it contained even one invented figure or quotation.

In a separate fictional merger exercise, Anthropic says Opus 5.5 built a financial model and an executive presentation in 63 minutes, compared with 93 minutes for Opus 5 on the same task, at half the cost. The company also described a demonstration in which the model tested four proposed explanations for a hypothetical retailer's slowing growth. It says the data supported none of them, although one could not be fully ruled out, and the model proposed another explanation involving the proximity of new stores to existing ones.

Sources

Companies in this story

Related stories

Explore

More articles