Claude Opus 5.5: what really changes (and how to use it to study)
AI news19 min read · 27 September 2026
On September 22, 2026, Anthropic released Claude Opus 5.5. On paper it’s the combination everyone wants: as good as the company’s top-end model, cheaper than the previous Opus, and faster. But there’s a gap between launch-day charts and what you’ll actually get out of it for your classes. Let’s sort it out: what was announced, who measured it, and how to use it without letting it do your thinking for you.
In short: Claude Opus 5.5 is the first model in the Claude 5.5 family. Anthropic says it performs at the level of Claude Fable 5.1 on most work, costs 40% less to run than Opus 5, and generates output more than 30% faster. On the API it costs $4 per million input tokens and $20 per million output tokens. In the Claude apps it’s paid-only: the free plan gives you Sonnet and Haiku, not Opus. For studying, the real lever isn’t the model, it’s how you use it: as a coach that asks you questions, not as a ghostwriter.
Visual: Anthropic.
What Anthropic announced on September 22
According to Anthropic’s official announcement, Opus 5.5 is “the first model in our new Claude 5.5 family.” The headline claims, all from the company itself:
- performance at the level of Claude Fable 5.1 “on most work”;
- 40% lower costs than Opus 5 on typical workloads at default settings;
- output generated more than 30% faster than Opus 5;
- pre-release testing by outside evaluators, including METR and Frontier Design.
Anthropic also says Claude Sonnet 5.5 and Claude Haiku 5.5 will follow “in the coming weeks.” One more piece of context: this is the company’s first release since CEO Dario Amodei published an essay in September calling for a slower pace of progress. In it he writes: “We must slow the pace at which we improve the capabilities of AI models.” Shipping a new model days later isn’t necessarily a contradiction (the essay is about pace, not stopping), but TechCrunch pointed out the timing.
On the technical side, the model page lists a 1-million-token context window, up to 128,000 output tokens, and a reliable knowledge cutoff of June 2026.
Pricing, plainly
Two worlds to keep apart: the API (for developers, billed per token) and the Claude apps (monthly subscription).
| Price per million tokens (API) | Opus 5.5 | Opus 5 |
|---|---|---|
| Input | $4 | $5 |
| Output | $20 | $25 |
| Cache reads | $0.20 | $0.50 |
| Cache writes | $5 | $6.25 |
There’s also a fast mode, up to 2.5x quicker, at $8 input and $40 output per million tokens, per Anthropic. Batch requests are half price, the documentation adds.
In the apps, the plans page is clear: the free plan includes Sonnet and Haiku, not Opus. Opus comes with Pro, listed at $17 a month billed annually or $20 billed monthly (before tax), then Max from $100 a month. With this launch, Anthropic says it’s raising five-hour usage limits on the Pro, Max, Team and Enterprise plans.
Effort levels: the setting that matters
This is the least visible change and the most useful one to understand. On Opus 5.5, the model’s thinking is always on and can no longer be switched off. What you control is effort: how much the model reasons before answering. Anthropic’s documentation lists five levels:
- low: the most efficient, for simple tasks;
- medium: balanced, and the default on Opus 5.5 (other Claude models default to high);
- high: as much thinking as the task needs;
- xhigh: for long coding or agentic tasks;
- max: no limit on token spending.
More effort means more thinking, so more time and more cost. It doesn’t always mean a better answer, as the charts show.
The benchmarks Anthropic published
One caveat before the numbers: all of these results are published by Anthropic, in its own announcement. They are self-reported measurements. Unless noted otherwise, Opus 5.5 scores use max effort. On Terminal-Bench, the GPT-6 Astra and GPT-5.6 Sol figures are the ones reported by OpenAI. And Anthropic itself warns that at this level of capability, benchmark margins have become “a less reliable guide to real-world differences,” adding that in its own use the gap with Fable 5.1 is narrower than the scores suggest.
Each chart plots score against cost per task (log scale), with one point per effort level.
Terminal-Bench 4.0: multi-step command-line work
Terminal-Bench 4.0
View the data
| Model | low | med | high | xhigh | max |
|---|---|---|---|---|---|
| Opus 5.5 | 38.5 · 1.29 $ | 57.6 · 2.94 $ | 64.2 · 3.88 $ | 66.4 · 7.35 $ | 64.8 · 11.24 $ |
| Fable 5.1 | 40.2 · 5.7 $ | 43.4 · 7.8 $ | 49.4 · 10.5 $ | 51.3 · 15.8 $ | 55.8 · 19.5 $ |
| Opus 5 | 28.5 · 4.25 $ | 41.2 · 7 $ | 47 · 10.64 $ | 50.6 · 13.48 $ | 52.3 · 15.83 $ |
| GPT-6 Astra | 49.7 · 4.95 $ | 53.9 · 6.15 $ | 57.9 · 7.21 $ | 57.6 · 7.48 $ | 56.7 · 10.35 $ |
| GPT-5.6 Sol | 7.9 · 1.46 $ | 20.9 · 2.69 $ | 26.1 · 4.12 $ | 28.5 · 5.39 $ | 37.3 · 7.89 $ |
This test measures how well a model handles complex, multi-step tasks in a terminal. Best scores reported: Opus 5.5 at 66.4% (at xhigh), GPT-6 Astra at 57.9%, Fable 5.1 at 55.8%, Opus 5 at 52.3% and GPT-5.6 Sol at 37.3%. The most telling point is elsewhere: at the default medium setting, Opus 5.5 scores 57.6% for $2.94 per attempt, versus 52.3% for Opus 5 at max, at $15.83. Another detail: at max, Opus 5.5 does worse (64.8%) than at xhigh.
FrontierCode v1.1: code you’d actually merge
FrontierCode v1.1 (main set)
View the data
| Model | low | med | high | xhigh | max |
|---|---|---|---|---|---|
| Opus 5.5 | 47.3 · 0.4 $ | 54.64 · 0.8 $ | 53.99 · 1.09 $ | 51.42 · 2.25 $ | 54.43 · 6.19 $ |
| Fable 5.1 | 52.8 · 2.47 $ | 50.91 · 3.28 $ | 50.34 · 5.27 $ | 48.73 · 9.27 $ | 50.28 · 12.82 $ |
| Opus 5 | 41.95 · 2.64 $ | 53.38 · 4.61 $ | 47.99 · 7.62 $ | 43.65 · 8.99 $ | 48.04 · 12.28 $ |
| GPT-6 Astra | 45.27 · 1.59 $ | 48.83 · 2.28 $ | 50.94 · 2.85 $ | 50.62 · 3.1 $ | 53.26 · 4.36 $ |
| GPT-5.6 Sol | 35.44 · 1.75 $ | 39.93 · 2.5 $ | 45.06 · 3.25 $ | 46.84 · 3.88 $ | 47.49 · 4.85 $ |
FrontierCode checks whether an agent’s code changes would be accepted into a real codebase. Opus 5.5 at medium scores 54.6% for about $0.80 per task, above GPT-6 Astra’s best (53.3% at max, $4.36). At max, Opus 5.5 scores 54.4% for $6.19: almost eight times the cost, with no gain. Fable 5.1 tops out at 50.3%, Opus 5 at 48.0%, GPT-5.6 Sol at 47.5%.
CursorBench 4.0: vague, multi-file requests
CursorBench 4.0
View the data
| Model | low | med | high | xhigh | max |
|---|---|---|---|---|---|
| Opus 5.5 | 43.7 · 1.18 $ | 52.5 · 2.9 $ | 56 · 3.97 $ | 56 · 6.99 $ | 57.8 · 13.43 $ |
| Fable 5.1 | 45.1 · 5.44 $ | 46.8 · 7.05 $ | 49.2 · 9.08 $ | 51.6 · 13.01 $ | 51.8 · 17.28 $ |
| Opus 5 | 40.7 · 4.87 $ | 43.3 · 6.94 $ | 44.7 · 9 $ | 46.1 · 11.43 $ | 46.6 · 11.95 $ |
| GPT-5.6 Sol | 24.6 · 0.87 $ | 31.1 · 1.77 $ | 35.7 · 2.85 $ | 37.7 · 4.4 $ | 41.7 · 8.23 $ |
This benchmark uses ambiguous tasks taken from real sessions in the Cursor editor. Opus 5.5 reaches 57.8% at max ($13.43) and 52.5% at medium ($2.90). Against it: Fable 5.1 at 51.8%, Opus 5 at 46.6%, GPT-5.6 Sol at 41.7% ($8.23). GPT-6 Astra isn’t on this chart.
GDPval-AA v2.1: office work across 44 occupations
GDPval-AA v2.1
View the data
| Model | low | med | high | xhigh | max |
|---|---|---|---|---|---|
| Opus 5.5 | 1224 · 0.21 $ | 1576 · 0.86 $ | 1692 · 1.54 $ | 1820 · 4.21 $ | 1846 · 8.92 $ |
| Fable 5.1 | 1450 · 1.41 $ | 1536 · 2.17 $ | 1617 · 3.43 $ | 1721 · 7.09 $ | 1735 · 9.59 $ |
| Opus 5 | 1294 · 0.58 $ | 1476 · 1.36 $ | 1581 · 3.03 $ | 1676 · 4.96 $ | 1708 · 6.76 $ |
| GPT-6 Astra | 1366 · 0.85 $ | 1468 · 1.82 $ | 1485 · 2.43 $ | 1516 · 3.04 $ | 1542 · 4.53 $ |
| GPT-5.6 Sol | 1289 · 0.27 $ | 1403 · 0.6 $ | 1480 · 1.11 $ | 1548 · 1.7 $ | 1588 · 2.81 $ |
This Artificial Analysis benchmark ranks models with an Elo score on real professional tasks. At max, Opus 5.5 reaches 1,846, ahead of Fable 5.1 (1,735), Opus 5 (1,708), GPT-5.6 Sol (1,588) and GPT-6 Astra (1,542). At medium it scores 1,576 for $0.86 per task, beating Astra at max for about a fifth of the cost. But at low it drops to 1,224, the lowest score on the whole chart.
The limits worth knowing
First caveat, and a big one: OpenAI released GPT-6 Sol and GPT-6 Luna the same day, about an hour later according to Simon Willison. GPT-6 Sol costs $2 input and $10 output, half the price of GPT-5.6 Sol. Anthropic’s charts compare Opus 5.5 with GPT-5.6 Sol, not with the newcomer.
Second caveat: max effort can go off the rails. Simon Willison, a developer and blogger known for his model tests, reports that his usual prompt (an SVG of a pelican riding a bicycle) failed twice at max: the model hit the 128,000-token output limit while still reasoning, after nearly 20 minutes and $2.56 per attempt. He now suspects that max is “effectively useless.” Fable 5.1 at max did answer.
Third caveat, flagged by Anthropic itself: Opus 5.5 “often suspects it is being evaluated,” which makes its real-world behavior harder to assess. Finally, the glowing testimonials in the announcement (GitHub, Stripe, Lovable and others) come from customers Anthropic picked for its launch page. They’re interesting, not independent.
Cheaper alternatives: Sonnet 5 and Haiku 4.5
You don’t need Opus to review a chapter. Anthropic’s documentation positions each model like this:
| Model | Positioning (per Anthropic) | API price (input / output) | In the free app |
|---|---|---|---|
| Fable 5.1 | Demanding reasoning, long-horizon work | $10 / $50 | No |
| Opus 5.5 | Long-running coding and knowledge work | $4 / $20 | No |
| Sonnet 5 | Best balance of speed and intelligence | $2 / $10 | Yes |
| Haiku 4.5 | The fastest | $1 / $5 | Yes |
Sonnet 5, released on June 30, is the default model on the Free and Pro plans, and Anthropic describes its performance as close to Opus 4.8. For explaining a lecture, generating a quiz or checking your reasoning, that’s plenty.
How to use it to study
The trap with a stronger model is handing it more of the work. Your brain only improves when it does the lifting. Here’s a method that keeps the AI in the coach’s seat.
1. Match the model to the task. Sonnet (free) for explanations, quizzes and summary sheets. Opus only for the heavy stuff: a multi-file coding project, a long document to analyze, a multi-step problem where Sonnet loses the thread.
2. Ask for questions, not answers. Paste your notes, then: “Ask me five questions of increasing difficulty on this chapter. Wait for my answer before correcting, and tell me exactly where my reasoning breaks down.” Anthropic’s education offering even highlights a “learning mode” that works like a tutor asking you questions.
3. Get guided through a problem, not handed the solution. In thermodynamics or algorithms, share your partial solution and ask only for the next hint. Still stuck? A second hint. The full solution comes last, if at all.
4. Teach it back. Explain the concept in your own words and ask the model to find the gaps, Feynman-style. That’s where you find out whether you understood it or just read it.
5. Check anything factual. A date, a formula, a reference: cross-check with your course or a reliable source. Even a good model gets things wrong.
6. Remember that generated text is marked. Opus 5.5 ships with Anthropic’s watermarking to comply with the EU AI Act. According to Anthropic, the watermark is invisible, carries no information about you, and a detection API is in private preview for eligible organizations, including educational ones. One more reason to write your assignments yourself and check your school’s rules on AI use.
Our take
Opus 5.5 is mostly good news for developers who pay per token: top-tier performance, a lighter bill, and a medium setting that’s often enough. For students, the change is more modest. On the free plan you don’t get it, and Sonnet 5 covers most study use cases. If you already pay for Pro, use it for long tasks, but don’t max out effort on principle: Anthropic’s own charts show max costing far more for zero, or even negative, gain on several tests. And OpenAI’s price cut the same day is a reminder that no ranking lasts long.
What it concretely changes for you: nothing decisive if you’re revising, a real gain if you code substantial projects, and a good excuse to get into the habit of choosing your model and effort level per task.
FAQ
Is Claude Opus 5.5 free?
Not in the apps. According to Anthropic’s plans page, the free plan includes Sonnet and Haiku, but not Opus. You need at least Pro ($17 a month billed annually or $20 monthly, before tax).
Opus 5.5 or Sonnet 5 for studying?
For explaining a lecture, building quizzes or checking your reasoning, Sonnet 5 is enough in the vast majority of cases. Opus 5.5 pulls ahead on long, complex tasks, like a multi-file coding project or analyzing a large document.
Should I set effort to max for the best answers?
No. On several benchmarks Anthropic published, max does no better than medium or xhigh, at a much higher cost. Start with the default and only raise it if answers are clearly too shallow.
Can I trust Anthropic’s benchmarks?
They’re self-reported by the company, using its own methods and settings. They show a trend, not an absolute truth. Anthropic itself writes that benchmark margins have become a less reliable guide to real-world differences.
Can a teacher tell if a text came from Claude?
Anthropic applies an invisible watermark to text from its recent models, and a detection API is in preview for certain organizations, including educational ones. The best answer is still the simplest: write your own work and use AI to understand and practice.
Further reading
- Studying with AI without cheating: the full method for keeping AI in a coaching role.
- New AI models in 2026: which one should you study with?: an overview of every major consumer model and its plans.
- The Step-by-step exercise skill: it guides you hint by hint instead of handing you the answer.
- The Feynman method skill: check you’ve understood something by explaining it in your own words.
Sources
- Introducing Claude Opus 5.5 — Anthropic · accessed 27 September 2026
- Claude Opus 5.5 (model page) — Claude Platform Docs · accessed 27 September 2026
- Models overview — Claude Platform Docs · accessed 27 September 2026
- Effort — Claude Platform Docs · accessed 27 September 2026
- Pricing (Claude plans) — Anthropic · accessed 27 September 2026
- Introducing Claude Sonnet 5 — Anthropic · accessed 27 September 2026
- How Claude’s text watermark works — Anthropic · accessed 27 September 2026
- Claude for Education — Anthropic · accessed 27 September 2026
- Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war — Simon Willison · accessed 27 September 2026
- Anthropic releases Opus 5.5 with lower prices and Fable-level performance — TechCrunch · accessed 27 September 2026
- Introducing GPT-6 Sol and Luna — OpenAI · accessed 27 September 2026
- We Must Pace the Frontier — Dario Amodei · accessed 27 September 2026
