
Anthropic: Drop 'think hard'
Anthropic says Claude Opus 5.5 at Medium effort matched or beat its predecessor at High. That is a vendor claim, so test it on your own evals.
Anthropic's guidance for Claude Opus 5.5 is to delete "think hard" prompts and start at Medium effort, because Anthropic reports Opus 5.5 at Medium matched or beat Claude Opus 5 at High on its coding and knowledge-work evaluations. The claim comes from the vendor, so teams should rerun it on their own evals.
The Weights Desk · 3 min read- Anthropic made Claude Opus 5.5 the default model on Pro, Max, Ultra and Team plans in Claude.
- Anthropic's advice is to delete "think hard" and "think carefully" from prompts, since thinking is always on and scales with the effort setting.
- Start at Medium effort and move to High only when Medium stalls.
- The headline result, Opus 5.5 at Medium matching or beating Opus 5 at High, is Anthropic's own and has no published harness we can point to.
- Anthropic says removing the phrases made replies start sooner with no clear drop in quality; no latency or token figure is given for that change.
Anthropic has made Claude Opus 5.5 the default model on Pro, Max, Ultra and Team plans in Claude, and its accompanying guide, dated November 12, 2026, tells users to stop writing "think hard" into prompts and to start at Medium effort. The supporting claim is that Opus 5.5 at Medium matched or beat Claude Opus 5 at High. That is a vendor result without a linked harness, so it is a hypothesis to test rather than a finding.
What Anthropic actually changed
The concrete instruction is to delete "think hard" and "think carefully" from prompts. Anthropic's guide says thinking is always on for Opus 5.5 and that the model decides how much to think from the request and the effort level. It adds that removing the phrases made replies start sooner, with no clear drop in quality. The guide gives no latency or token figures for that specific change.
Medium is the default, High is the escalation
Anthropic says Medium is the default effort for Opus 5.5 and advises starting there, moving to High when Medium stalls. Users change it by clicking the model name next to the send button and choosing Effort. The justification is that Opus 5.5 at Medium matched or beat Opus 5 at High on coding and knowledge-work evaluations. Anthropic also quotes Factory saying Medium matched High while using 20 to 25 percent fewer output tokens.
The evidence is vendor-supplied
None of the comparisons come with a benchmark name, score or test harness a reader can rerun. The guide reports that 16 of 18 research reports passed across effort settings and that a merger model took 63 minutes on Opus 5.5 against 93 on Opus 5 at half the cost. These are useful anecdotes, but they are Anthropic's own selections, and one customer quote is not a controlled comparison.
Price makes the effort setting an economic choice
Anthropic lists Opus 5.5 at $4 input and $20 output per million tokens, and Sonnet 5.5 at $2 and $10. It also says that at higher effort settings, cost per task with Sonnet 5.5 gets close to Opus 5.5. The Decoder words this more strongly, saying Sonnet ends up paying about as much. Anthropic's wording is the one to rely on, and the crossover point depends on the task.
Verdict: adopt the default, verify before you bake it in
Removing "think hard" and starting at Medium is a low-risk change and worth trying, because it is cheap to reverse and Anthropic's guidance points the same way. Do not treat the parity claim as settled. Run your own eval set at Medium and High, record latency, output tokens and pass rate, and keep High for the tasks where Medium demonstrably fails. Watch for a published harness that would let outsiders check Anthropic's numbers.
- What should teams change in their Claude Opus 5.5 prompts?
- Anthropic says to delete "think hard" and "think carefully". Thinking is always on for Opus 5.5 and adjusts to the request and the effort setting. Anthropic says replies then start sooner with no clear drop in quality.
- Which effort level should be the starting point?
- Medium, which Anthropic says is the default for Opus 5.5. Its guidance is to move up to High when Medium stalls, rather than starting high by habit.
- Is the Medium-versus-High result independently verified?
- No. It comes from Anthropic's own evaluations and a customer quote it cites. No harness or per-benchmark scores are linked in the guide, so a team should reproduce it on its own tasks before changing production settings.