Claude Opus 5.5 Guide: Drop 'Think Hard', Start at Medium
Anthropic's daily-driver guide for Opus 5.5 says to drop 'think hard', start at Medium effort and state the finished goal in one message.
Anthropic's daily-driver guide for Opus 5.5 says to drop 'think hard', start at Medium effort and state the finished goal in one message.
Introduction
On November 12, 2026, Anthropic published "Using Claude Opus 5.5 as your daily driver for knowledge work," a guide by Aleksandra Todorova that tells users which prompting habits to drop. Anthropic says Opus 5.5 is now the default model in Claude on Pro, Max, Ultra, and Team plans, and recommends it as the model to leave selected for research, analysis, and producing memos, reports, spreadsheets, and decks. This review is about the guidance rather than the model's launch: what Anthropic tells people to change, what evidence it offers, and where its supporting API documentation adds nuance.
Feature Overview
Drop "think hard." Anthropic says to delete phrases such as "think hard" and "think carefully" from prompts. According to the guide, thinking is always on for Opus 5.5, and the model decides how much to think based on the request and the effort level. In Anthropic's testing, removing those lines made replies start sooner with no clear drop in quality.
Start at Medium effort. Medium is the default. Anthropic says that on its coding and knowledge-work evaluations Opus 5.5 at Medium matched or beat Opus 5 at High, and that higher settings think longer and cost more without always producing better work. The advice is to move up to High only when Medium stalls. To change it, click the model name next to the send button and choose Effort. Anthropic adds a general rule for any model, which is to try a higher effort first if Claude underdelivers, and move up a model only if Claude had the context, clearly tried, and still got it wrong.
State the goal once. Instead of instructing Claude step by step, Anthropic says to describe the goal, the finished output, and when you want Claude to stop and ask. It says Opus 5.5 carries long, multipart work through to the end better than Opus 5 did. Its demo prompt for a retail revenue question names the goal and the final deliverable but does not say which file to check for which theory.
Attach, ask for the finished file, name what you dislike. Anthropic says to share a chart or diagram itself rather than retyping its numbers, because workarounds built for earlier models may no longer be needed. It says to ask for the spreadsheet or document directly rather than an outline, since Opus 5.5 can often reach the finished file in one pass. For design work, it says a vague instruction like "avoid a generic look" tends to swap one default style for another, while naming patterns such as a cream background or pill-shaped buttons works better.
Connect files and systems. Anthropic calls connectors one of the most useful steps, because Claude can fetch what a task needs instead of working only from pasted text. On Team and Enterprise plans an owner has to turn a connector on before members can use it. The guide also suggests pasting brand or writing guidelines, and warns to expect a few rounds before output matches.
Usability Analysis
The guide's model table gives each tier a job, with prices per million input and output tokens.
| Model | Anthropic's suggested use | Price (input / output) |
|---|---|---|
| Haiku 5.5 | Quick lookups and high-volume work | $0.10 / $0.50 |
| Sonnet 5.5 | Quick, well-scoped tasks and polished documents from a clear brief | $2 / $10 |
| Opus 5.5 | Most knowledge work, including finished docs, decks and spreadsheets that need judgment | $4 / $20 |
| Fable 5.2 | The most ambitious, open-ended technical work | $8 / $40 |
Anthropic notes that Haiku 5.5's prices apply to prompts up to 100k tokens, and that Pro plans and standard seats on Team and seat-based Enterprise plans do not include Fable 5.2 in their limits outside limited-time launch offers, so it runs on usage credits.
Anthropic's argument is that a more capable model can finish a task with less work, offsetting much of its higher per-token price. It says Sonnet 5.5's cost per task gets close to Opus 5.5's at higher-effort settings, so users pushing Sonnet to Extra high or Max should consider Opus 5.5. The API documentation adds nuance. On a SWE-bench Pro subset, Anthropic reports Opus 5.5 at low effort solved about the same share of tasks as Sonnet 5.5 at its default for about 7% more per solved task. It tells readers to price candidates in cost per completed task on their own traffic.
Anthropic's internal examples are specific but small. In a research test, an automated grader checked every figure and quote against a saved web snapshot and failed any report with a single invented one, and 16 of 18 reports passed across effort settings. In a test merger analysis between two fictional HR software companies, Opus 5.5 finished in 63 minutes against 93 for Opus 5, at half the cost. Anthropic also says that even at its lowest effort setting Opus 5.5 read dense chart values more accurately than Opus 5 at its highest.
Effort and Cost in Anthropic's Docs
Anthropic's API documentation publishes an effort sweep for Opus 5.5 on a SWE-bench Pro subset, noting the scores are not comparable to the public leaderboard.
| Effort | Tasks solved | Cost per task |
|---|---|---|
| Low | 87.4% | $0.11 |
| Medium (default) | 92.8% | $0.21 |
| High | 95.3% | $0.29 |
| Extra high (single run) | 96.7% | $0.74 |
Read together with the guide, the numbers show why Medium is the starting point and also that it is a trade-off. Anthropic says Medium scored about 2.5 points lower than High for about 70% of the cost per task. The documentation also advises sweeping two or three effort levels in separate sessions, because changing effort mid-session invalidates the cache and distorts the comparison.
Pros and Cons
Pros: the advice is concrete and easy to test, the model-selection rule (raise effort before switching models) is simple, and Anthropic backs the cost argument with published effort and cost-per-task figures.
Cons: the evidence is Anthropic's own internal testing plus early-tester quotes, Medium trades a few points of accuracy for cost in the published sweep, and connector-dependent advice requires an owner to enable connectors on Team and Enterprise plans.
Outlook
The guide frames each new model as a reason to retest what you hand to Claude and to drop habits built for older ones. That framing suggests prompt libraries copied from earlier model generations deserve an audit, especially ones that encode reasoning-depth phrases or step-by-step choreography. The documentation's emphasis on cost per completed task over per-token price points the same way. Teams can verify the claims cheaply by running one multi-file task at Medium with a single goal-describing message and comparing the result with their usual prompts.
Conclusion
This is a practical vendor guide with enough published numbers to test its main claims. The most transferable advice is to remove reasoning-depth phrases, start at Medium, state the finished deliverable once, and attach source material directly. Treat the benchmark comparisons as Anthropic's own and verify on your workload. It is most useful for people who use Claude for research, analysis and document production, and who carry prompt habits from older models.
Editor's Verdict
Claude Opus 5.5 Guide: Drop 'Think Hard', Start at Medium earns a solid recommendation within the Claude space.
The strongest case for paying attention: concrete habits that are quick to test on a single real task, such as removing reasoning-depth phrases and starting at Medium. That alone raises the bar for what readers should expect in this space. Reinforcing that is a simple escalation rule, which is to raise effort first and move up a model only if Claude had the context and still failed — practical value rather than just headline appeal. The broader signal worth registering is straightforward: prompt hygiene becomes a cost lever, because Anthropic ties deleting reasoning-depth phrases to faster replies with no clear drop in quality. On the other side of the ledger, one constraint is real rather than a marketing footnote: evidence comes from Anthropic's own internal tests and early-tester quotes, and the guide names no independent evaluation. It should factor into any serious decision. Layered on top of that, Medium effort trades a few accuracy points for lower cost in Anthropic's own coding sweep, so it is not a free upgrade — which narrows the set of teams for whom this is an obvious yes.
For people who use Claude daily for research, analysis and document work, and teams maintaining prompt libraries written for older models, this guide is worth testing against a real task this week.
Pros
- Concrete habits that are quick to test on a single real task, such as removing reasoning-depth phrases and starting at Medium
- Simple escalation rule, which is to raise effort first and move up a model only if Claude had the context and still failed
- Published effort sweep with solved rates and cost per task, so the Medium default can be weighed against its cost
- Plan coverage stated plainly, with Opus 5.5 as the default on Pro, Max, Ultra, and Team plans
Cons
- Evidence comes from Anthropic's own internal tests and early-tester quotes, and the guide names no independent evaluation
- Medium effort trades a few accuracy points for lower cost in Anthropic's own coding sweep, so it is not a free upgrade
- Connector-based advice depends on an owner enabling connectors on Team and Enterprise plans
References
Comments0
Key Features
1. Anthropic says to delete phrases like "think hard" because thinking is always on for Opus 5.5 and scales with the request and the effort level. 2. Start at Medium effort, the default, which Anthropic says matched or beat Opus 5 at High on its coding and knowledge-work evaluations; move to High only when Medium stalls. 3. Describe the goal, the finished output and when to stop and ask in a single first message instead of instructing step by step. 4. Attach charts and diagrams as files, ask directly for the finished document or spreadsheet, and name specific unwanted design styles. 5. Use connectors so Claude can fetch files and systems itself, and use cost per completed task rather than per-token price to compare Sonnet, Opus and Fable tiers.
Key Insights
- Prompt hygiene becomes a cost lever, because Anthropic ties deleting reasoning-depth phrases to faster replies with no clear drop in quality.
- Effort level replaces prompt wording as the control for reasoning depth, since Anthropic says thinking is always on and the model scales it to the request and the setting.
- Anthropic says Opus 5.5 at Medium matched or beat Opus 5 at High on its evaluations, yet its own API docs show Medium about 2.5 points below High on a coding subset for about 70% of the cost.
- The Sonnet-versus-Opus advice is conditional, because the docs report Opus 5.5 at low effort solving about the same share of tasks as Sonnet 5.5 at default for about 7% more per solved task.
- Goal-first prompting changes how the work is split, with Anthropic saying Opus 5.5 carries long, multipart work through to the end better than Opus 5 did.
- Vision gains change input habits, because Anthropic says chart values read at the lowest effort beat Opus 5 at its highest, so retyping numbers may no longer help.
- Connectors are the step Anthropic stresses most for tailored output, but on Team and Enterprise plans an owner must enable them first.
- Anthropic's documentation pushes buyers to price cost per completed task on their own traffic and to test effort levels in separate sessions.
Was this review helpful?
Share
Related AI Reviews
Claude Fable 5.2 Paints With Code, Not Diffusion
Anthropic staff demos show Fable 5.2 painting by writing code for a simulated paint engine, with roughly 700,000 strokes in one photorealistic piece.
Claude Personal Finance: Read-Only Bank Access in Beta
Personal Finance in Claude links bank and card accounts read-only for Pro, Max and Ultra users in the US, rolling out in beta from this week.
Anthropic FlashProve: Zero-Knowledge Proofs of AI Inference
Anthropic's open-source FlashProve prototype proves a reply came from a committed model without exposing weights, at about 110 tokens/s on one H100.
Anthropic S-1/A Warns RSP Threshold May Be Met 'Imminently'
Revised S-1 says internal estimates for future models may 'imminently' meet its RSP acceleration threshold; no current model is designated above it.
