Claude Sonnet 5.5
Claude Sonnet 5.5 is the middle model of Anthropic's 5.5 family, released on Sep 28, 2026. It costs US$ 2 per million input tokens and US$ 10 per million output, the same as Sonnet 5 and GPT-6 Sol, with a 1 million token window and no surcharge. In Anthropic's figures, it beats Opus 5.5 on terminal work and automation and stays close on the rest, at half the price per token. The advantage disappears at max effort: in three of four tests with published cost, an Opus 5.5 effort level scores higher for less money per task.
Claude Sonnet 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-sonnet-5-5- Release
- Sep 28, 2026
- Input
- US$ 2 per million tokens
- Output
- US$ 10 per million tokens
- Cache read
- US$ 0.20 per million (10% of input)
- Cache write
- US$ 2.50 (5 minutes) or US$ 4 (1 hour)
- Batch API
- US$ 1 input and US$ 5 output
- Context window
- 1 million tokens, no surcharge
- Max output
- 128 thousand tokens (300 thousand in batch, beta)
- Reliable knowledge
- through June 2026
- Reasoning
- adaptive; cannot be turned off, the minimum is
between_tools - Default effort
- high on the API; medium in the app and in Claude Code
- Where it runs
- Anthropic API, Amazon Bedrock, Google Cloud, Microsoft Foundry, Claude Platform on AWS and GitHub Copilot paid plans
- Retirement
- not before Sep 28, 2027
Where Claude Sonnet 5.5 does well, where it falls behind and who it is for
Where it does well
- Long terminal work: 70.6% on Terminal-Bench 4.0, against 66.4% for Opus 5.5, which costs twice as much per token.
- Automation: 44.7% on AutomationBench, against 42.5% for Opus 5.5 and 32.0% for GPT-6 Sol.
- Office work: 1,844 points on GDPval-AA v2.1, two below Opus 5.5.
- Cost at the usual levels: from low to xhigh effort, it is cheaper per task than Opus 5.5 on all four tests with published cost.
- A generational jump: it beats Sonnet 5 on all nine tests in the table, at the same price.
Where it falls behind
- Max effort: on CursorBench 4.0, its best is 55.5% for US$ 9.67 per task; Opus 5.5 at high scores 56.0% for US$ 3.97.
- Heavy coding: 81.3% on SWE-Bench Pro, against 89.9% for Opus 5.5, the widest gap in the table.
- Terminal at medium effort: 28.8% on Terminal-Bench 4.0, against 57.6% for Opus 5.5 at the same level.
- Saying it does not know: per the system card, it is slightly more likely to state a wrong answer than to admit it does not know.
- Migration: five changes return errors in code written for Sonnet 5.
Who it makes sense for
- Well-scoped tasks at volume: bug fixes, code generation, documents, spreadsheets and decks.
- Terminal agents and workflow automation, run at high effort or above.
- Apps that repeat a long instruction, with cache reads at US$ 0.20 and 1 million tokens with no surcharge.
- Anyone comparing it with GPT-6 Sol: the price per token is identical, and above 272 thousand tokens per request only OpenAI charges more.
When to pick another model
- Open-ended, long work that depends on judgment: Anthropic itself points to Opus 5.5.
- When you are tempted to raise it to xhigh or max: run the same tasks on Opus 5.5 at medium or high first.
- Classifying and summarizing at maximum volume: Haiku 4.5 costs half.
- High-risk security research: those requests fall back to Sonnet 5, because of the cybersecurity safeguards.
What Anthropic published at launch
Sonnet 5.5 shipped six days after Opus 5.5, as its fast, cheap complement. Anthropic describes it as strong on well-scoped tasks, bug fixing and work with documents, spreadsheets and decks, and leaves Opus with open-ended work that needs judgment over a long stretch.
Against Opus 5.5, it wins two tests and loses seven, most of them narrowly. On the terminal win, the system card reports a standard error of 2.5 points for Sonnet 5.5 and 2.6 for Opus 5.5: the 4.2 point gap is a trend, not a comfortable lead.
| Sonnet 5.5 Anthropic | Opus 5.5 Anthropic | Sonnet 5 Anthropic | GPT-6 Sol OpenAI | |
|---|---|---|---|---|
| Terminal-Bench 4.0 long terminal work |
70.6% | 66.4% | 10.3% | not published |
| AutomationBench automation |
44.7% | 42.5% | 10.7% | 32.0% |
| GDPval-AA v2.1 office work |
1,844 | 1,846 | 1,449 | 1,487 |
| OSWorld 2.1 computer use |
80.1% | 81.8% | 57.0% | not published |
| FrontierCode 1.1 coding |
52.1% | 54.4% | 42.4% | 49.3% |
| CursorBench 4.0 coding |
55.5% | 57.8% | 34.1% | not published |
| SWE-Bench Pro coding |
81.3% | 89.9% | 63.2% | not published |
| Humanity's Last Exam hard knowledge, with tools |
64.5% | 67.7% | 54.9% | not published |
| Chartography chart reading |
61.6% | 64.4% | 15.6% | 53.6% |
Figures from the Claude Sonnet 5.5 announcement and system card. The dot marks the best in each row. On FrontierCode, Sonnet 5.5 scored 52.1% at xhigh and 46.2% at max; we show the higher one. On AutomationBench, Opus 5.5 appears as remeasured with the fallback model on (40.0% at its own launch).
Opus 5.5, Sonnet 5.5 or Haiku 4.5: which to choose
Anthropic's documentation tells most workloads to start with Opus 5.5, which takes agents that run alone for hours and large refactors. Sonnet 5.5 comes in when speed and cost matter: everyday coding, data analysis, content creation and agents with tools. Haiku 4.5 is the lowest latency, lowest price option, for real time, high volume and sub-agents.
Haiku 4.5 costs half of Sonnet 5.5 per token, but it has a 200 thousand token window and reliable knowledge only through February 2025. Anthropic promised Haiku 5.5 for the coming weeks, with no date or price, and the documentation only guarantees Haiku 4.5 until Oct 15, 2026.
Between Sonnet and Opus, the rule comes from the cost charts: from low to xhigh, Sonnet 5.5 is the cheaper one per task. If the task calls for the maximum, test Opus 5.5 before raising Sonnet.
| Opus 5.5 | Sonnet 5.5 | Haiku 4.5 | |
|---|---|---|---|
| Input | US$ 4 | US$ 2 | US$ 1 |
| Output | US$ 20 | US$ 10 | US$ 5 |
| Cache read | US$ 0.20 | US$ 0.20 | US$ 0.10 |
| Context window | 1 million | 1 million | 200 thousand |
| Max output | 128 thousand | 128 thousand | 64 thousand |
| Reliable knowledge | Jun 2026 | Jun 2026 | Feb 2025 |
| Latency, per Anthropic | moderate | fast | fastest |
Anthropic documentation, pricing and models overview, checked on Oct 1, 2026.
The max effort trap
There are five effort levels, from low to max. More effort means more reasoning, more tokens and more time. And the levels were recalibrated: Sonnet 5.5's medium is not Sonnet 5's medium.
Anthropic's own cost-per-task charts show where the math flips. On CursorBench 4.0, Sonnet 5.5 climbs from 47.8% at high to 55.5% at max, but goes from 37,391 to 271,920 tokens per task. That comes to US$ 9.67 per task, against US$ 3.97 for Opus 5.5 at high, which scores 56.0%. On FrontierCode, Sonnet's max costs US$ 20.78 per task, and Opus's US$ 6.19.
The exception is Terminal-Bench 4.0, where Sonnet at max scores 70.6% for US$ 12.54 per attempt, above any Opus level. Time counts too: on the system card's health professional questions, an answer takes 12 to 26 seconds from low to xhigh and about 160 at max, to gain between 1 and 1.6 points.
The documentation's starting point: high for general use, medium for agents with well-defined tasks, low or medium for chat. Then measure.
What breaks when coming from Sonnet 5
Swapping the identifier is not enough. There are five changes that return errors, summarized below, and a sixth that raises no error at all: the text the model writes between tool calls now comes back inside thinking blocks, empty by default. A screen that showed that text to the user goes blank, with no alarm.
- Forced tool use:
tool_choiceset toanyortoolreturns a 400 error. Useautowith strict mode or structured outputs. - Thinking off:
disabledno longer exists. The minimum isbetween_tools, accepted up to high effort. - Rewritten history: on accounts created on or after Aug 31, 2026, resending reasoning after changing the system prompt, the tools or an earlier message returns a 400 error.
- Computer use: the
computer_20251124tool is not accepted on the Anthropic API or Google Cloud. - Advisor model: Opus 4.8, Opus 4.7 and Sonnet 5 are no longer accepted in that role.
Examples of using Claude Sonnet 5.5
An agent takes each ticket, reproduces the bug in the terminal, fixes it and opens the review request. It is the short, well-scoped task profile Anthropic describes for Sonnet 5.5, at high effort, where it still costs less than Opus.
Each conversation becomes a record with the customer's name, topic and intent. The fixed format comes from a tool in strict mode or structured outputs, because the forced tool use that worked on Sonnet 5 now returns a 400 error.
The month's reports become a spreadsheet and a deck. On GDPval-AA, which measures this kind of deliverable, Sonnet 5.5 sits two points below Opus 5.5, at half the price per token.
Where Claude Sonnet 5.5 stands out and where it falls behind, by the numbers
Claude Sonnet 5.5 has published results on 23 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Jan 1, 2026 and Dec 27, 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 23 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 95.6% | #3 / 73 | #3 / 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 |
| CritPt | 31.4% | #5 / 76 | #5 / 112 | 32.3% GPT-5.6 Sol |
| SciCode | 61.0% | #4 / 74 (tied with 1) | #4 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 88.8% | #7 / 64 | #7 / 86 | 93.7% GPT-6 Astra |
| WebDev Arena | 1,786 | #3 / 67 | #3 / 82 | 1,820 Claude Opus 5.5 |
| SimpleQA Verified | 46.5% | #31 / 59 | #36 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 59.0% | #8 / 58 (tied with 2) | #9 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 65.0% | #4 / 58 | #4 / 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 | 80.5% | #8 / 54 | #8 / 67 | 100.0% GPT-6.1 Sol |
| Apex Agents | 75.5% | #2 / 48 | #2 / 50 | 82.4% Gemini 4 Argon |
| FrontierCode | 52.1% | #5 / 43 | #5 / 43 | 54.6% Claude Opus 5.5 |
| Hle | 46.1% | #4 / 13 | #4 / 39 | 54.8% GPT-6 Astra |
| Gdp Pdf | 31.4% | #3 / 36 | #3 / 37 | 35.0% GPT-6.1 Sol |
| Furniture Assembly | 75.0% | #4 / 32 | #4 / 34 | 83.3% Claude Opus 5.5 |
| FrontierSWE | 61.9% | #3 / 20 | #3 / 20 | 65.5% GPT-6 Astra |
| CursorBench | 55.5% | #2 / 16 | #2 / 16 | 57.8% Claude Opus 5.5 |
| Weirdml V3 | 19.9% | #4 / 10 | #4 / 11 | 42.2% GPT-6 Astra |
| Frontiermath Erdos | 2.9% | #3 / 7 | #3 / 7 | 2.9% GPT-6 Astra |
| Spireclimb | 1.3% | #6 / 7 | #6 / 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 | 2.7% | #6 / 7 | #6 / 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 Sonnet 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 |
|---|---|---|---|---|---|
| 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 |
| Claude Sonnet 5.5 | Anthropic | 164.45 | 1,537 | 77.8 | Sep 28, 2026 |
| GPT-6 Sol | OpenAI | 162.55 | n/a | 79.3 | Sep 22, 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 Sonnet 5.5 or Gemini 4 Argon?
Gemini 4 Argon is the model from another lab closest to Claude Sonnet 5.5 in capability. Across the 17 tests both took, this is the score:
| Test | Claude Sonnet 5.5 | Gemini 4 Argon |
|---|---|---|
| SimpleQA Verified | 46.5% | 76.0% |
| CritPt | 31.4% | 27.1% |
| FrontierSWE | 61.9% | 55.0% |
| Gdp Pdf | 31.4% | 28.4% |
| FrontierCode | 52.1% | 47.6% |
| Apex Agents | 75.5% | 82.4% |
| WebDev Arena | 1,786 | 1,680 |
| CursorBench | 55.5% | 52.9% |
| Furniture Assembly | 75.0% | 72.5% |
| Frontiermath Tiers 1 3 V2 | 88.8% | 86.0% |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
What Claude Sonnet 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 Sonnet 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 Sonnet 5.5
How much does Claude Sonnet 5.5 cost?
US$ 2 per million input tokens and US$ 10 per million output. Cache reads cost US$ 0.20, cache writes US$ 2.50 (5 minutes) or US$ 4 (1 hour), and batch is US$ 1 and US$ 5. The price holds across the whole 1 million token window.
Sonnet 5.5 or Opus 5.5?
For short, high-volume tasks, Sonnet 5.5, which costs half per token and is cheaper per task from low to xhigh effort. For long, open-ended work, or when Sonnet would need max, test Opus 5.5: on CursorBench 4.0 it scores 56.0% for US$ 3.97 at high, while Sonnet's best costs US$ 9.67 to reach 55.5%. The independent scores are at the top of this page.
Sonnet 5.5 or Haiku 4.5?
Haiku 4.5 costs US$ 1 and US$ 5 per million tokens and is the fastest in the lineup, but it has a 200 thousand token window and reliable knowledge only through February 2025. For classifying and summarizing at volume, it does the job. Anthropic promised Haiku 5.5 for the coming weeks, with no date.
Do I need to change my code to use Sonnet 5.5?
Yes, if it uses forced tool use, turns thinking off, rewrites the history before resending, uses the old computer use tool or Sonnet 5 as an advisor. Those cases return errors. Also redo your effort level test, because the levels were recalibrated.
Where does Claude Sonnet 5.5 rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, Claude Sonnet 5.5 ranks #6 out of 267 models, with an ECI of 164.45. 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 Sonnet 5.5 better than Gemini 4 Argon?
Across the 17 tests both models took and Epoch AI compiled, Claude Sonnet 5.5 comes out ahead on 11, behind on 4 and tied on 2. 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 Sonnet 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 Sonnet 5.5 are: ProofBench (#1 of 64); Apex Agents (#2 of 48); GPQA Diamond (#3 of 73). 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 Sonnet 5.5 announcement
- Anthropic, Claude Sonnet 5.5 system card
- Anthropic, what's new in Claude Sonnet 5.5 (documentation)
- Anthropic, pricing (documentation)
- Anthropic, models overview (documentation)
- Anthropic, choosing a model (documentation)
- OpenAI, API pricing (GPT-6 Sol)
- GitHub Changelog, Claude Sonnet 5.5 in GitHub Copilot
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: Sonnet 5.5: benchmarks, pricing and the real cost per task
← See the full ranking