If you opened Claude this week and your prompts felt slower than they should, there is a reason. Anthropic now treats Claude Opus 5.5 as the default model on Pro, Max, Ultra and Team plans, and on 12 November 2026 it published a guide telling people which old prompting habits to drop.
The short version: stop writing “think hard”, stay on Medium effort, and ask for the finished file in one message instead of drafting step by step.
This guide covers every habit Anthropic listed, what to write instead (with copy-ready prompts), how to change effort on web, desktop and the API, what each plan gets, the vendor numbers behind the advice, and when another model is the better pick.
- What changed: Opus 5.5 is the default model in Claude on Pro, Max, Ultra and Team, and Anthropic now calls it the “daily driver” for research, writing, spreadsheets and decks.
- Thinking is always on: you cannot switch it off on Opus 5.5. Effort (Low to Max) is the only dial.
- Default effort is Medium, one step lower than Opus 5. Anthropic says Opus 5.5 on Medium matched Opus 5 on High in its tests (vendor-reported).
- Drop these habits: “think hard” lines, starting on High, drip-feeding instructions, asking for a draft first, retyping charts, vague “no generic design” rules, and working without connectors.
- API:
claude-opus-5-5, $4 input / $20 output per million tokens, 1M context, 128K max output. Sendingthinking: disabledreturns a 400 error. - Free plan: still Sonnet and Haiku only, no Opus.
What Anthropic Announced and When
Opus 5.5 itself is not new. Anthropic released it on 22 September 2026 with lower prices than Opus 5 and higher usage limits for subscribers.
What is new is the positioning. On 12 November 2026, Anthropic published “Using Claude Opus 5.5 as your daily driver for knowledge work” on claude.com. The Decoder and other outlets picked it up on 14 November.
The guide says Opus 5.5 is the default on Pro, Max, Ultra and Team, and recommends keeping it selected for most work. It then lists the prompting habits that slow it down.
| Date (2026) | Event | Source |
|---|---|---|
| 22 Sep | Claude Opus 5.5 released: $4/$20 per million tokens, higher five-hour limits on paid plans | Anthropic launch page |
| 30 Oct | Max 20x renamed Ultra ($200/month); Max 5x becomes Max ($100/month) | Anthropic help centre, BuildMVPFast report |
| 12 Nov | “Daily driver” guide published, with Opus 5.5 named the default and 7-8 prompting changes | claude.com resources |
| 14 Nov | The Decoder and DEV Community cover the guide | The Decoder, DEV |
Which Plans Get Opus 5.5 by Default
Every paid individual and team plan includes Opus 5.5 inside its normal usage limits. The Free plan does not.
| Plan | Price | Opus 5.5 | Fable 5.2 |
|---|---|---|---|
| Free | $0 | No (Sonnet and Haiku only) | No |
| Pro | $20/month ($17/month billed yearly) | Yes, default | Only via paid usage credits |
| Max | $100/month | Yes, default | Yes, up to 50% of weekly limits |
| Ultra | $200/month | Yes, default | Yes |
| Team Standard seat | $25/seat ($20 yearly) | Yes, default | Via credits |
| Team Premium seat | $125/seat ($100 yearly) | Yes, default | Yes, up to 50% of weekly limits |
Note: Anthropic says pricier models use up your plan limits faster. Picking Opus over Sonnet costs nothing extra in dollars on a subscription, but you can hit the five-hour or weekly cap sooner.
Enterprise admins can switch off a model or an effort level for certain roles. If you do not see Opus 5.5 or the Effort menu at work, ask your admin first.
8 Prompting Habits to Drop (and What to Do Instead)
Anthropic’s guide lists seven changes, and The Decoder splits them into eight. Here is each one, the reason behind it, and a replacement you can paste.
1. Stop writing “think hard” or “think carefully”
These phrases date back about a year and a half, to the first reasoning models. On Opus 5.5 thinking is always on, and the model decides how deep to go based on the task and effort level.
Anthropic says that in its tests, replies without these lines started faster with no clear drop in quality. The developer docs give the same advice for system prompts.
- Old: “Think step by step and think very carefully before you answer.”
- New: just state the task. If the answer is too shallow, raise effort instead.
2. Stop starting on High effort
Medium is the default for Opus 5.5. Anthropic recommends starting there and moving to High only when Medium stalls.
Its internal tests say Opus 5.5 on Medium matched Opus 5 on High. An early tester at Factory reported the same match while using 20 to 25% fewer output tokens (vendor-reported).
3. Stop defaulting to Sonnet “to save money”
Many people keep Sonnet selected and switch to Opus only for hard work. Anthropic argues Opus 5.5 often finishes in fewer steps, which can offset its higher per-token price.
The guide adds that running Sonnet 5.5 at Extra high or Max effort costs about as much per task as Opus 5.5. If you find yourself pushing Sonnet that far, switch models.
4. Stop giving instructions one at a time
Drip-feeding a chat (“first read this”, “now summarise”, “now make a table”) wastes turns. Describe the end result in the first message.
Anthropic’s checklist for that first message:
- The goal: what decision or question the work should answer.
- The finished output: format, length, audience.
- When to stop and ask: which choices Claude should check with you instead of guessing.
5. Stop asking for a draft or outline first
Opus 5.5 often goes from request to finished document, spreadsheet or deck in one pass. Ask for the final file directly and review that.
Claude can export documents to Word, PDF or Google Docs and decks to PowerPoint, PDF or Google Slides, according to the guide.
6. Stop retyping charts and tables
Attach the chart, screenshot or diagram as an image or file. Anthropic says Opus 5.5 is its best Opus model for vision, and even at its lowest effort it read chart values more accurately than Opus 5 at its highest.
Tip: for dense technical drawings, upload a higher-resolution image. The developer docs say technical drawings benefit most from extra resolution.
7. Stop saying “avoid a generic design”
Vague bans just swap one default look for another. Name the exact patterns you do not want.
- Old: “Make it look modern and not like generic AI output.”
- New: “Do not use a cream or off-white background, italic accent words in headlines, numbered 01/02/03 section labels, monospace labels or pill-shaped buttons.”
Check which defaults the first result uses, then add those to your list. Sharing your own brand or writing guide helps too, and Anthropic says to expect a few revision rounds.
8. Stop working without connectors
Connect Claude to the files and systems your work lives in (Google Drive, email, CRM and so on). It can then pull the context it needs instead of relying on what you paste.
On Team and Enterprise plans, an owner has to enable connectors first. Our Claude Projects guide covers the other way to give Claude standing context: project knowledge files.
Before and After: Rewriting a Real Prompt
Here is a typical “old style” prompt for a sales analysis, and the same request written the way Anthropic now recommends.
Before (old habits):
Think very carefully. First, read the attached sales file and tell me
what you see. Then we'll make an outline for a report.
Make the charts look nice and not generic.
After (Opus 5.5 style):
Goal: find out why Q3 revenue growth slowed. Leadership has four theories
(listed below); test each against the attached sales spreadsheets and
the two PDFs.
Output: a finished 2-page memo that leads with the answer, plus 3 charts.
Use our style guide (attached). No cream backgrounds, no pill buttons,
no italic accent words in headings.
Ask me before: dropping any store from the data, or making assumptions
about missing months.
The second version gives the goal, the finished output and the check-in rules in one go. In Anthropic’s own demo, a single prompt like this led Claude to reject all four theories and find a fifth explanation (new stores opening too close to old ones), backed by a regression.
How to Change Effort in Claude (Web, Desktop, Mobile)
The effort menu sits in the same place on every Claude app, per Anthropic’s help centre.
- Open the model menu. Click the model name next to the send button.
- Pick the model. Choose Opus 5.5, or click “More models” for the full list.
- Open Effort. Click “Effort” in the same menu. The recommended level is marked “Default”.
- Choose a level. Low, Medium, High, Extra high or Max. The change applies to Claude’s next reply.
| Effort | Use it for | Effect on limits |
|---|---|---|
| Low | Quick lookups, rewording, short answers | Stretches usage furthest |
| Medium (default) | Most research, writing and analysis | Balanced |
| High | When Medium misses details or stalls | Uses limits faster |
| Extra high | Long coding and agent tasks | Much faster usage |
| Max | The hardest one-off problems | Fastest usage, slowest replies |
There is no “Thinking” toggle for Opus 5.5. The help centre confirms thinking is always on for Opus 5.5, Fable 5.2 and Fable 5.1, so lowering effort is the only way to reduce it.
For Developers: Effort and Thinking in the API
The API follows the same rules. Set effort with output_config.effort; if you leave it out, Opus 5.5 runs at medium, one level lower than Opus 5’s default of high.
import anthropic
client = anthropic.Anthropic()
response = client.messages.create(
model="claude-opus-5-5",
max_tokens=16000,
messages=[{"role": "user", "content": "Review this contract for inconsistencies."}],
output_config={"effort": "high"},
)
for block in response.content:
if block.type == "text":
print(block.text)
Key details from Anthropic’s docs:
- Valid values:
low,medium,high,xhigh,max. Do not passadaptive; that is a thinking mode, not an effort level. - No off switch:
thinking: {"type": "disabled"}returns a 400 error on Opus 5.5 at every effort level. - Leave room: thinking counts toward
max_tokens. Anthropic says 128,000 works well for long agentic coding turns. - Don’t assume block order: the first content block may be a thinking block, so check each block’s type.
- Caching: changing top-level effort between requests can invalidate the prompt cache. Use per-message effort (beta) to switch mid-conversation.
- Reasoning extraction: prompts that push the model to copy its full reasoning into the reply can trigger a
reasoning_extractionrefusal. Ask for a short explanation instead.
Warning: if you migrated from Opus 5 without setting effort, your requests quietly dropped from high to medium. Run a quick sweep on your own test set before assuming quality is the same.
For a deeper look at how effort changes your bill across Claude, OpenAI and Gemini, see our reasoning effort and API cost guide.
Opus 5.5 API specs at a glance
| Spec | Value |
|---|---|
| Model ID | claude-opus-5-5 (Bedrock: anthropic.claude-opus-5-5) |
| Context window | 1M tokens |
| Max output | 128K tokens (300K on Batches with a beta header) |
| Input / output price | $4 / $20 per million tokens |
| Cache read / 5-min write / 1-hour write | $0.20 / $5 / $8 per million tokens |
| Fast mode | $8 / $40, up to 2.5x faster |
| Batch API | 50% off input and output |
| Knowledge cutoff | June 2026 |
| Platforms | Claude API, AWS, Google Cloud, Microsoft Foundry |
The Numbers Behind the Advice
All of these come from Anthropic or its early testers. None are independent benchmarks, so treat them as claims to check on your own work.
| Claim | Figure | Who says so |
|---|---|---|
| Medium vs Opus 5 High | Matched or beat | Anthropic internal tests |
| Output tokens at Medium vs Opus 5 High | 20-25% fewer | Factory engineer (early tester) |
| Research reports with every figure and quote correct | 16 of 18 | Anthropic, automated grader on a saved web snapshot |
| Fictional merger model + exec deck | 63 min vs 93 min for Opus 5, at half the cost | Anthropic |
| Output speed vs Opus 5 | Over 30% faster | Anthropic developer docs |
| Terminal-Bench 4.0 | 66.4% (Opus 5: 52.3%) | Anthropic launch page, max effort |
| GDPval-AA v2.1 (Elo) | 1846 (Opus 5: 1708) | Anthropic launch page |
Anthropic itself notes on the launch page that benchmark margins are a weak guide to real-world differences. The Medium-vs-High claim is the easiest one to test yourself: run the same task at both levels and compare.
Opus 5.5 vs Sonnet 5.5 vs Haiku 5.5 vs Fable 5.2
Anthropic’s own rule of thumb: try a higher effort level before moving up a model. Only switch up if Claude had the context, clearly tried, and still got it wrong.
| Model | Best for | API price (in/out per 1M) | Default effort |
|---|---|---|---|
| Haiku 5.5 | Quick lookups, summaries, sorting, high volume | $0.10 / $0.50 ($0.50 / $2.50 over 100K) | Medium |
| Sonnet 5.5 | Well-scoped tasks with a clear brief, polished docs and slides | $2 / $10 | High |
| Opus 5.5 | Most knowledge work that needs judgement | $4 / $20 | Medium |
| Fable 5.2 | Long, open-ended research and large coding jobs with subagents | $8 / $40 | High |
Prices are from Anthropic’s 12 November guide and model pages. If you mostly code, our best Claude model for coding comparison goes deeper, and the Claude Haiku 5.5 guide covers the cheap end.
Against rivals, Anthropic’s launch table puts Opus 5.5 ahead of GPT-6 Astra on Terminal-Bench 4.0 (66.4% vs 57.9%) but behind it on Terminal-Bench-Science (58.7% vs 64.6%) and AutomationBench (40.0% vs 41.4%). Those are vendor-run numbers too.
Common Problems and Fixes
| Problem | Likely cause | Fix |
|---|---|---|
| Replies feel slow to start | Old “think hard” lines in your prompt, custom instructions or project | Delete them from your personal preferences in Settings and from any project instructions |
| Answers feel shallower than Opus 5 | Effort dropped from High to Medium by default | Raise to High for that chat |
| Hitting usage limits sooner | Extra high or Max left on | Go back to Medium; use Haiku for quick jobs |
| No Effort option visible | Older model selected, or admin restriction | Select Opus 5.5; on Enterprise, ask your admin |
| API returns 400 on thinking | thinking: {"type": "disabled"} sent to Opus 5.5 |
Remove the field and lower effort instead |
| API answer starts with an empty block | Thinking block with omitted display comes first | Loop over blocks and read only type == "text" |
Refusal with reasoning_extraction |
Prompt asks the model to dump its reasoning | Ask for a brief summary of how it got the answer |
| Design output still looks “AI generic” | Vague style ban | List the exact patterns to avoid and attach your style guide |
Limits and What Is Still Unclear
- Vendor numbers only: every speed, quality and cost figure above comes from Anthropic or partners it quotes. No independent group has re-run the Medium-vs-High comparison yet.
- Limits are not in dollars: on subscriptions, a heavier model drains your five-hour and weekly caps faster. Anthropic does not publish exact message counts.
- Free users are left out: the pricing page lists only Sonnet and Haiku for Free accounts.
- Pricing page lag: as of 15 November, claude.com/pricing still names Fable 5.1 as the newest Fable model, while the guide and docs list Fable 5.2.
- Privacy: connectors give Claude access to your files and accounts. Only connect what the task needs, and keep untrusted content out of searchable sources, as Anthropic’s own docs advise.
If you are new to Claude in general, what Claude is good for is a friendlier starting point than this guide.
My Take
This is the first time Anthropic has openly told paying users that their prompting habits are the problem. Most of the advice is sound and easy to try today.
The habit I would drop first is the “think hard” line hidden in custom instructions. Many people pasted it in a year ago and forgot it, and it is pure waste on a model that always thinks.
The money argument needs more care. “Opus finishes in fewer steps” is plausible, but it is Anthropic’s claim about its own pricier model. On a subscription, the real cost is how fast you hit your cap, so watch Settings > Usage for a week before you decide.
My practical setup: Opus 5.5 on Medium for daily work, High when a reply misses something, Haiku for quick rewrites, and Fable only for long research or big code jobs.
FAQ
Is Claude Opus 5.5 the default model now?
Yes, on Pro, Max, Ultra and Team plans. Anthropic confirmed this in its 12 November 2026 guide. Free accounts still use Sonnet and Haiku.
Can I turn off thinking on Opus 5.5?
No. Thinking is always on in the Claude apps and the API. Lower the effort level to reduce how much it thinks.
Should I still write “think step by step”?
Anthropic says no. In its tests, removing these phrases made replies start faster with no clear quality loss.
What effort level should I use?
Start on Medium, the default. Move to High if the answer misses details, and keep Extra high and Max for long coding or very hard problems.
Does Opus 5.5 cost more than Sonnet on a Pro plan?
Not in dollars, since the subscription price is fixed. It does use your usage limits faster than Sonnet or Haiku.
How much does Opus 5.5 cost on the API?
$4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 and a 50% discount on the Batch API.
Why do I get a 400 error when I disable thinking?
Opus 5.5 rejects thinking: {"type": "disabled"} at every effort level. Remove that field and set output_config.effort instead.
Is Opus 5.5 better than Fable 5.2?
Not for everything. Anthropic recommends Fable 5.2 for the most demanding long-running research and coding, and Opus 5.5 for most everyday knowledge work at half the API price.
Sources
- Claude resources, “Using Claude Opus 5.5 as your daily driver for knowledge work” by Aleksandra Todorova (12 Nov 2026)
- Anthropic, “Introducing Claude Opus 5.5” launch page (22 Sep 2026)
- Claude developer docs: Effort (output_config.effort, per-model defaults)
- Claude developer docs: Prompting Claude Opus 5.5
- Claude developer docs: Claude Opus 5.5 model overview (context, output, pricing, model IDs)
- Claude Help Centre, “Change the model, effort and thinking settings” (article 8664678)
- Claude pricing page, claude.com/pricing (checked 15 Nov 2026)
- The Decoder, “Anthropic says outdated prompting habits are slowing down Claude Opus 5.5” (14 Nov 2026)
- DEV Community, “Anthropic positions Claude Opus 5.5 as the default daily driver for knowledge work” (14 Nov 2026)
- Notebookcheck, “Claude Opus 5.5 is included with Pro and Free still misses out” (23 Sep 2026)
- BuildMVPFast, “Claude Max Plan Changes: Max 20x Is Now Ultra” (4 Nov 2026)




