Claude Opus 5.5
Claude Opus 5.5 has been Anthropic's main model since 22 September 2026. It costs US$ 4 per million input tokens and US$ 20 per million output, rereads the cache for US$ 0.20 and works with up to 1 million tokens of context. In the table Anthropic itself published, it wins seven of nine tests and loses two to GPT-6 Astra. It is the model for long, autonomous work, such as coding agents and multi-step office tasks; for simple volume work, Sonnet 5.5 delivers nearly the same at half the price per token.
Claude Opus 5.5 in official numbers
Data published by the maker in its documentation and announcement. The source of each item is at the end of the page.
- Maker
- Anthropic
- API identifier
claude-opus-5-5- Release
- Sep 22, 2026
- Input
- US$ 4 per million tokens
- Output
- US$ 20 per million tokens
- Cache read
- US$ 0.20 per million (5% of input)
- Cache write
- US$ 5 (5 minutes) or US$ 8 (1 hour)
- Batch API
- US$ 2 input and US$ 10 output
- Fast mode
- US$ 8 and US$ 40, Anthropic API only
- Context window
- 1 million tokens, no surcharge
- Max output
- 128 thousand tokens (300 thousand in batch, beta)
- Reliable knowledge
- through June 2026
- Reasoning
- adaptive, always on; default effort medium on the API
- Where it runs
- Anthropic API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS
- Retirement
- not before Sep 22, 2027
Where Claude Opus 5.5 does well, where it falls behind and who it is for
Where it does well
- Long terminal work: 66.4% on Terminal-Bench 4.0 in Anthropic's table, against 57.9% for GPT-6 Astra and 55.8% for Fable 5.1.
- Coding: leads FrontierCode 1.1 (54.4%) and CursorBench 4.0 (57.8%) in the same comparison.
- Office work: 1,846 points on GDPval-AA v2.1, the highest score in the table, with Sonnet 5.5 two points behind.
- Cache bill: rereading costs US$ 0.20 per million, 60% less than Opus 5. For apps with long instructions, it is the line that weighs most on the invoice.
- Agent behavior: according to Anthropic, it tried to get around limits about 85% less than Opus 5 in its automated audit.
Where it falls behind
- Science in the terminal: 58.7% on Terminal-Bench-Science, against 64.6% for GPT-6 Astra. It is the only clear loss in the maker's own table.
- Automation: 40.0% against 41.4% for GPT-6 Astra on the launch AutomationBench, a small but recorded gap.
- Price per token: twice Sonnet 5.5, which comes within two points of it on GDPval-AA and 1.7 points on OSWorld 2.1.
- Fast mode: doubles the price, to US$ 8 and US$ 40, and does not exist outside the Anthropic API.
- Reasoning always on: thinking cannot be turned off, and Anthropic rates its latency as moderate, against fast for Sonnet 5.5.
Who it makes sense for
- Agents that work alone for hours: coding, code migration, tracking down bugs in large systems.
- Multi-step office tasks, such as reports, spreadsheets and decks built from many documents.
- Assistants with long, repeated instructions (company handbook, catalog, policy), where the cheap cache cuts the cost.
- Anyone who does not yet know which model to use: it is the starting point Anthropic's own documentation recommends.
When to pick another model
- High volume of simple tasks, such as classifying, summarizing and extracting: Sonnet 5.5 or Haiku 4.5 do it for much less.
- Long scientific research inside the terminal, where GPT-6 Astra scored higher even in Anthropic's table.
- Answers that must come out instantly, such as live customer chat with short messages: Sonnet 5.5 is faster.
- Reasoning at the very edge: Anthropic points to Fable 5.1 when Opus 5.5 at high effort still falls short.
What Anthropic published at launch
Opus 5.5 opened Anthropic's 5.5 family on 22 September 2026 with an unusual move: it got better and cheaper at the same time. The price fell 20% from Opus 5, from US$ 5 and US$ 25 to US$ 4 and US$ 20, and cache reads fell 60%, from US$ 0.50 to US$ 0.20.
The company says it generates output more than 30% faster than Opus 5 and runs 40% cheaper on typical workloads, counting the steps it saves. That second figure is the maker's estimate, hard to check from the outside, because it depends on what counts as a typical workload.
The table below is the one from the announcement, against Fable 5.1 and Opus 5, both in-house, and GPT-6 Astra, from OpenAI. Anthropic chose the tests; it shows two losses, which is already more honest than the average launch.
| Opus 5.5 Anthropic | Fable 5.1 Anthropic | Opus 5 Anthropic | GPT-6 Astra OpenAI | |
|---|---|---|---|---|
| Terminal-Bench 4.0 long terminal work |
66.4% | 55.8% | 52.3% | 57.9% |
| FrontierCode 1.1 coding |
54.4% | 50.3% | 48.0% | 53.3% |
| CursorBench 4.0 coding |
57.8% | 51.8% | 46.6% | not measured |
| GDPval-AA v2.1 office work |
1,846 | 1,735 | 1,708 | 1,542 |
| AutomationBench automation |
40.0% | 31.4% | 26.9% | 41.4% |
| Humanity's Last Exam hard knowledge |
67.7% | 65.6% | 63.6% | 57.2% |
| Terminal-Bench-Science 0.1 science in the terminal |
58.7% | 52.6% | 29.0% | 64.6% |
| OSWorld 2.1 computer use |
81.8% | 80.7% | 74.0% | not measured |
| Chartography chart reading |
89.0% | 88.4% | 83.4% | not measured |
Figures from the Claude Opus 5.5 announcement. The dot marks the best in each row. "Not measured" cells are tests Anthropic did not publish for GPT-6 Astra.
Opus 5.5, Sonnet 5.5 or Fable 5.1: choosing within Anthropic
Choosing within the family is harder than it looks, because the cheaper model is not always cheaper per task. Sonnet 5.5 costs half per token (US$ 2 and US$ 10) and comes close to Opus on several tests. But in the cost-per-task charts Anthropic published with Sonnet, Opus 5.5 at high effort scores 56.0% on CursorBench for US$ 3.97 per task, while Sonnet 5.5 at its best scores 55.5% for US$ 9.67.
At the other end, Fable 5.1 costs US$ 10 and US$ 50 and is what Anthropic recommends for reasoning at the edge and very long autonomous work, when Opus 5.5 at high effort still falls short. In the Opus 5.5 launch table, however, Fable 5.1 does not win any of the nine tests.
In practice: start with Opus 5.5 at medium effort, the API default. Drop to Sonnet 5.5 if the task is short and high volume. Move up to Fable 5.1 only after measuring that Opus does not get there.
| Opus 5.5 | Sonnet 5.5 | Fable 5.1 | Haiku 4.5 | |
|---|---|---|---|---|
| Input | US$ 4 | US$ 2 | US$ 10 | US$ 1 |
| Output | US$ 20 | US$ 10 | US$ 50 | US$ 5 |
| Cache read | US$ 0.20 | US$ 0.20 | US$ 0.25 | US$ 0.10 |
| Context window | 1 million | 1 million | 1 million | 200 thousand |
| Latency, per Anthropic | moderate | fast | slower | fastest |
Anthropic documentation, pricing and models overview, checked on Oct 1, 2026.
How to test before switching
Benchmarks answer whether a model is capable; they do not answer whether it solves your case at your price. The test that decides takes an afternoon: pick ten real tasks, the kind your team does every week, and run them on Opus 5.5 and on the model you use today.
Compare two columns, accuracy and cost per task, not price per token. Turn on caching if your app repeats a long instruction, because that is where Opus 5.5 got cheaper. And leave effort at the default before raising it: high effort costs more and does not always improve the result.
Examples of using Claude Opus 5.5
An agent with access to the repository and the terminal converts modules, runs the tests and fixes what broke, for hours, without someone approving every step. It is the kind of task where Opus 5.5 leads Terminal-Bench 4.0.
A month of spreadsheets, PDFs and emails becomes a report with tables and a deck. The 1 million token window fits the whole material, and GDPval-AA measures exactly this kind of professional deliverable.
The handbook goes in as a fixed instruction and stays cached. Every question rereads it for US$ 0.20 per million tokens instead of paying full input, which is the line that weighs most as usage grows.
Where Claude Opus 5.5 stands out and where it falls behind, by the numbers
Claude Opus 5.5 has published results on 31 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Dec 26, 2025 and Dec 21, 2026 that took the same test. Against the whole archive, which includes 2023 models, a recent release lands near the top of almost everything, and that says nothing.
Where it stands out
Tests where it sits in the top third of its own generation.
Where it falls behind
Tests where it sits in the bottom half of its own generation.
See all 31 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 90.6% | #33 / 73 | #35 / 173 | 96.4% Gemini 4 Argon |
| AIME 2024 and 2025 (OTIS) | 100.0% | #1 / 75 (tied with 11) | #1 / 168 | 100.0% GPT-5.5 |
| Dtbench | 98.9% | #1 / 64 | #1 / 148 | 98.9% Claude Opus 5.5 |
| Lmca | 68.23 | #1 / 64 | #1 / 121 | 68.23 Claude Opus 5.5 |
| CritPt | 31.7% | #2 / 76 (tied with 2) | #2 / 112 | 32.3% GPT-5.6 Sol |
| SciCode | 66.9% | #1 / 74 | #1 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 91.2% | #3 / 64 | #3 / 86 | 93.7% GPT-6 Astra |
| ARC-AGI 2 | 93.3% | #3 / 43 | #3 / 83 | 95.0% GPT-6 Astra |
| ALE-Bench | 2,147 | #5 / 57 | #5 / 82 | 2,951 GPT-6 Astra |
| WebDev Arena | 1,820 | #1 / 67 | #1 / 82 | 1,820 Claude Opus 5.5 |
| ARC-AGI | 98.5% | #1 / 43 (tied with 4) | #1 / 81 | 98.5% Claude Fable 5 |
| SimpleQA Verified | 72.2% | #5 / 59 | #5 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 67.0% | #4 / 58 | #4 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 71.0% | #3 / 58 | #3 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 100.0% | #1 / 64 (tied with 2) | #1 / 74 | 100.0% Claude Fable 5.1 |
| Frontiermath Tier 4 V2 | 95.0% | #3 / 54 | #3 / 67 | 100.0% GPT-6.1 Sol |
| Vending Bench 2 | 9,235 | #9 / 43 | #9 / 57 | 15,515 GPT-6 Astra |
| Apex Agents | 73.5% | #3 / 48 | #3 / 50 | 82.4% Gemini 4 Argon |
| FrontierCode | 54.6% | #1 / 43 | #1 / 43 | 54.6% Claude Opus 5.5 |
| Gdp Pdf | 30.6% | #6 / 36 | #6 / 37 | 35.0% GPT-6.1 Sol |
| Furniture Assembly | 83.3% | #1 / 32 | #1 / 34 | 83.3% Claude Opus 5.5 |
| Blueprint Bench 2 | 51.2% | #2 / 29 | #2 / 31 | 54.4% Gemini 4 Argon |
| Ebr Bench | 71.4% | #2 / 22 | #2 / 27 | 76.2% GPT-6 Astra |
| FrontierSWE | 62.3% | #2 / 20 | #2 / 20 | 65.5% GPT-6 Astra |
| CursorBench | 57.8% | #1 / 16 | #1 / 16 | 57.8% Claude Opus 5.5 |
| Rli | 21.3% | #1 / 9 | #1 / 15 | 21.3% Claude Opus 5.5 |
| MirrorCode | 77.4% | #1 / 9 | #1 / 9 | 77.4% Claude Opus 5.5 |
| Frontiermath Erdos | 2.9% | #2 / 7 | #2 / 7 | 2.9% GPT-6 Astra |
| Spireclimb | 3.3% | #4 / 7 | #4 / 7 | 25.7% GPT-6 Astra |
| Spireclimb Sts1 A20 Heart | 0.0% | #4 / 7 (tied with 3) | #4 / 7 | 20.0% GPT-6.1 Sol |
| Spireclimb Sts2 A10 | 6.7% | #4 / 7 | #4 / 7 | 34.7% GPT-6 Astra |
For each test, the model's best effort level counts, as in the ranking. Source of the scores: Epoch AI, under CC BY 4.0.
What to compare Claude Opus 5.5 with
The models from other labs closest in capability and released in the same period, at most two per lab. Click any of them to see its page.
| Model | Lab | ECI | ELO | LiveBench | Release |
|---|---|---|---|---|---|
| Claude Opus 5.5 | Anthropic | 166.82 | 1,584 | 83.2 | Sep 22, 2026 |
| GPT-6 Astra | OpenAI | 166.41 | n/a | 82.2 | Sep 3, 2026 |
| GPT-6.1 Sol | OpenAI | 166.32 | n/a | 81.6 | Sep 29, 2026 |
| Gemini 4 Argon | Google DeepMind | 164.56 | 1,608 | n/a | Sep 30, 2026 |
| Kimi K3 | Moonshot | 157.42 | 1,528 | 79.2 | Jul 16, 2026 |
| Gemini 3.7 Flash | Google DeepMind | 157.22 | 1,551 | 78.8 | Aug 13, 2026 |
| Grok Build 0.1 | xAI | no ECI | n/a | 67.8 | May 29, 2026 |
Claude Opus 5.5 or GPT-6 Astra?
GPT-6 Astra is the model from another lab closest to Claude Opus 5.5 in capability. Across the 30 tests both took, this is the score:
| Test | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|
| Spireclimb Sts1 A20 Heart | 0.0% | 16.7% |
| Spireclimb | 3.3% | 25.7% |
| Spireclimb Sts2 A10 | 6.7% | 34.7% |
| Vending Bench 2 | 9,235 | 15,515 |
| MirrorCode | 77.4% | 46.7% |
| ALE-Bench | 2,147 | 2,951 |
| SciCode | 66.9% | 56.5% |
| Game puzzles | 71.0% | 84.0% |
| Apex Agents | 73.5% | 64.7% |
| Chess puzzles | 67.0% | 76.0% |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
What Claude Opus 5.5 costs to deliver each task
Price per token is what is on the price sheet; cost per task is what lands on the invoice, because it depends on how many tokens the model spends. Each point is an effort level, published by whoever ran the test and compiled by Epoch AI. The further up and to the left, the more score per dollar.
CursorBench
The score scale does not start at zero.
Other models from Anthropic
The models from the same lab closest in time, newest first.
| Model | Release | ECI | ELO | LiveBench | Access |
|---|---|---|---|---|---|
| Claude Haiku 5.5 | Oct 7, 2026 | 158.98 | n/a | 72.1 | API access |
| Claude Sonnet 5.5 | Sep 28, 2026 | 164.45 | 1,537 | 77.8 | API access |
| Claude Opus 5.5 | Sep 22, 2026 | 166.82 | 1,584 | 83.2 | API access |
| Claude Fable 5.1 | Sep 1, 2026 | 164.65 | 1,586 | 83.4 | API access |
| Claude Opus 5 | Jul 24, 2026 | 162.60 | 1,573 | 80.1 | API access |
| Claude Sonnet 5 | Jun 30, 2026 | 156.36 | 1,491 | 76.0 | API access |
| Claude Fable 5 | Jun 9, 2026 | 162.13 | 1,542 | 83.0 | API access |
| Claude Opus 4.8 | May 28, 2026 | 158.26 | 1,507 | 76.2 | API access |
| Claude Opus 4.7 | Apr 16, 2026 | 156.32 | 1,532 | 76.5 | API access |
Have you used Claude Opus 5.5? Rate it
Users' opinion, not measurement: the rating is kept apart from the ranking and does not move anyone's position. One vote per browser, no sign-up, and you can change your rating later.
What people ask about Claude Opus 5.5
How much does Claude Opus 5.5 cost?
US$ 4 per million input tokens and US$ 20 per million output. Cache reads cost US$ 0.20, cache writes US$ 5 (5 minutes) or US$ 8 (1 hour), and batch processing is half price. Fast mode, on the Anthropic API only, costs US$ 8 and US$ 40.
Is Opus 5.5 better than GPT-6 Astra?
In the table Anthropic published, Opus 5.5 wins seven of nine tests, including the coding and long terminal work ones, and loses two to GPT-6 Astra: science in the terminal (58.7% against 64.6%) and automation (40.0% against 41.4%). The independent scores are at the top of this page, updated by the sources.
Is it worth paying twice the price of Sonnet 5.5?
It depends on the task. For short, high-volume tasks, Sonnet 5.5 delivers nearly the same at half the price per token. For long, hard tasks, Opus 5.5 can be cheaper per task: on CursorBench it scores 56.0% for US$ 3.97 at high effort, while Sonnet 5.5 needs US$ 9.67 to reach 55.5%.
Where can I use Opus 5.5?
On the Anthropic API (identifier claude-opus-5-5), Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS, as well as in the Claude apps. Anthropic guarantees it will not be retired before 22 September 2027.
Where does Claude Opus 5.5 rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, Claude Opus 5.5 ranks #1 out of 267 models, with an ECI of 166.82. The ECI is a composite index that combines several hard benchmarks into a single scale. The position changes when a new model comes out, and the page is refreshed six times a day.
Is Claude Opus 5.5 better than GPT-6 Astra?
Across the 30 tests both models took and Epoch AI compiled, Claude Opus 5.5 comes out ahead on 11, behind on 16 and tied on 3. It is a test-by-test count with no weighting: a win in math counts the same as a win in coding. To choose between the two, look at the tests that resemble your own use.
What is Claude Opus 5.5 best at?
Against models of the same generation (released up to 9 months before and up to 3 months after it), the best relative results of Claude Opus 5.5 are: SciCode (#1 of 74); WebDev Arena (#1 of 67); Dtbench (#1 of 64). The yardstick is its position among the models of that period that took the same test.
Where do the numbers on this page come from?
From the same three sources as the ranking, all openly licensed: Epoch AI (ECI and individual tests, CC BY 4.0), Arena (human-preference ELO, CC BY 4.0) and LiveBench (overall average, CC BY-SA 4.0). ROO3 does not measure any model. When the page shows a figure published by the maker itself, it appears in the written analysis, labeled as the maker's, and never enters the ranking.
Where each number comes from
- Anthropic, Claude Opus 5.5 announcement
- Anthropic, models overview (documentation)
- Anthropic, pricing (documentation)
- Anthropic, Claude Sonnet 5.5 announcement (cost-per-task charts including Opus 5.5)
Epoch AI, "AI Benchmarking Hub". Published at epoch.ai. Data under the CC BY 4.0 licence, reordered by roo3.co.
The Arena (LMArena) leaderboard, from the public lmarena-ai/leaderboard-dataset on Hugging Face, under the CC BY 4.0 licence. We are not affiliated with the Arena.
LiveBench, by the LiveBench team, published at livebench.ai under the CC BY-SA 4.0 license. The overall average is recomputed by roo3.co with the same formula as the website (average of the categories), and those numbers stay under CC BY-SA 4.0. We are not affiliated with LiveBench.
Read also: Opus 5.5: the expensive one got better and cheaper at the same time
← See the full ranking