Claude Sonnet 5
Claude Sonnet 5 shipped on Jun 30, 2026 at a launch price of US$ 2 per million input tokens and US$ 10 per million output, which became its permanent price in August. In Anthropic's figures, it came close to Opus 4.8 at less than half the price per token. It is now legacy: Sonnet 5.5, from Sep 28, 2026, costs the same and beats it on all eight tests in the launch table, but requires code changes. Sonnet 5 will not be retired before Jun 30, 2027.
Claude Sonnet 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- Release
- Jun 30, 2026
- Status
- legacy; the documentation says to migrate to Sonnet 5.5
- Input
- US$ 2 per million tokens (same as Sonnet 5.5)
- 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
- Fast mode
- not available
- Context window
- 1 million tokens, no surcharge
- Max output
- 128 thousand tokens (300 thousand in batch, beta)
- Reliable knowledge
- through January 2026
- Reasoning
- adaptive, on by default; turns off with
disabledat any effort level - Default effort
- high on the API and in Claude Code
- Cache minimum
- 1,024 tokens (512 on Sonnet 5.5)
- Where it runs
- Anthropic API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS
- Capacity
- does not support Priority Tier
- Retirement
- not before Jun 30, 2027
Where Claude Sonnet 5 does well, where it falls behind and who it is for
Where it does well
- Office work: 1,618 on GDPval-AA v2, a statistical tie with Opus 4.8 (1,615), which cost US$ 5 and US$ 25.
- Terminal coding: 80.4% on Terminal-Bench 2.1, against 67.0% for Sonnet 4.6 and 82.7% for Opus 4.8.
- Price: US$ 2 and US$ 10, below Sonnet 4.6's US$ 3 and US$ 15, with cache reads at US$ 0.20.
- Flexible on the API: it turns thinking off at any effort level and accepts forced tool use and the old computer use tool. Sonnet 5.5 rejects all three.
- Agent safety: per the system card, it resists prompt injection better than Sonnet 4.6 and hallucinates and flatters far less.
Where it falls behind
- Against its successor: 10.3% on Terminal-Bench 4.0, against 70.6% for Sonnet 5.5, which costs the same.
- Cost per task: on CursorBench 4.0, its best is 34.1% for US$ 7.17 per task; Sonnet 5.5 at low effort scores 35.8% for US$ 0.50.
- Automation: 13.5% on AutomationBench, below Gemini 3.5 Flash's 14.5% in its own system card.
- Features left behind: caching only from 1,024 tokens, no Priority Tier, no task budgets and no mid-conversation system messages.
- Over-refusal: in Claude Code cybersecurity tests, it refuses malicious requests more reliably than Sonnet 4.6, but also refuses more legitimate ones.
Who it makes sense for
- Code that turns thinking off, forces a tool or uses
computer_20251124and has not been updated yet: all of that returns an error on Sonnet 5.5. - Fallback model: it is where Sonnet 5.5 sends the higher-risk cybersecurity requests it refuses, billed at Sonnet 5 rates.
- Structured extraction with a tuned prompt: according to Anthropic, it reads prompts literally.
- Transition period: at the same price, you can run Sonnet 5 and Sonnet 5.5 side by side until at least Jun 30, 2027.
When to pick another model
- New projects: Sonnet 5.5 costs the same, wins all eight tests in the launch table and generates output more than 30% faster, according to Anthropic.
- Tasks that need recent knowledge: its reliable cutoff is January 2026; Sonnet 5.5's is June 2026.
- Classifying and summarizing at maximum volume: Haiku 4.5 costs half, US$ 1 and US$ 5.
- Advisor for a Sonnet 5.5 executor in the advisor tool: the request returns an error.
What Anthropic published at launch
Sonnet 5 arrived on Jun 30, 2026 as the most agentic Sonnet up to then, able to plan and use a browser and a terminal on its own. It became the default model on the Free and Pro plans.
In the announcement table, Opus 4.8 appears only for reference and leads five of the six rows; on GDPval-AA v2 the two tie (1,618 against 1,615). The GPT-5.5 column comes from the system card.
| Sonnet 5 Anthropic | Sonnet 4.6 Anthropic | Opus 4.8 Anthropic | GPT-5.5 OpenAI | |
|---|---|---|---|---|
| SWE-bench Pro coding |
63.2% | 58.1% | 69.2% | 58.6% |
| Terminal-Bench 2.1 terminal coding |
80.4% | 67.0% | 82.7% | 83.4% |
| Humanity's Last Exam without tools |
43.2% | 34.6% | 49.8% | 41.4% |
| Humanity's Last Exam with tools |
57.4% | 46.8% | 57.9% | 52.2% |
| OSWorld-Verified computer use |
81.2% | 78.5% | 83.4% | 78.7% |
| GDPval-AA v2 office work, Elo |
1,618 | 1,395 | 1,615 | 1,509 |
| FrontierCode v1 coding |
38.8% | 15.1% | not published | 25.5% |
Sonnet 5, Sonnet 4.6 and Opus 4.8: table from the Claude Sonnet 5 announcement. GPT-5.5 and FrontierCode: table 8.1.A of the system card (max effort, averaged over 5 trials). The dot marks the best in each row. GPT-5.5 on Terminal-Bench 2.1 was measured with the Codex CLI.
The launch price that stuck
Sonnet 5 debuted at an introductory price of US$ 2 and US$ 10 per million tokens, valid through Aug 31, 2026. From Sep 1, 2026, the planned price was US$ 3 and US$ 15, the same as Sonnet 4.6. On Aug 10, 2026, Anthropic dropped the increase and made the introductory price permanent.
Two details for the bill. The cost charts in the announcement still use US$ 3 and US$ 15, so the real cost is lower than they show. And the new tokenizer turns the same text into 1.0 to 1.35 times as many tokens as on Sonnet 4.6: the price per token fell, but the cost per text falls less.
| Sonnet 5 legacy | Sonnet 5.5 current | Opus 5.5 current | Haiku 4.5 current | |
|---|---|---|---|---|
| Input | US$ 2 | US$ 2 | US$ 4 | US$ 1 |
| Output | US$ 10 | US$ 10 | US$ 20 | US$ 5 |
| Cache read | US$ 0.20 | US$ 0.20 | US$ 0.20 | US$ 0.10 |
| Cache write 5 minutes |
US$ 2.50 | US$ 2.50 | US$ 5 | US$ 1.25 |
| Batch input and output |
US$ 1 and US$ 5 | US$ 1 and US$ 5 | US$ 2 and US$ 10 | US$ 0.50 and US$ 2.50 |
| Reliable knowledge | Jan 2026 | Jun 2026 | Jun 2026 | Feb 2025 |
| Retirement not before |
Jun 30, 2027 | Sep 28, 2027 | Sep 22, 2027 | Oct 15, 2026 |
Anthropic documentation: pricing, model pages and model deprecations, checked on Oct 1, 2026.
Is Claude Sonnet 5 still worth using in 2026?
As your main model, no. Sonnet 5.5 shipped 90 days later at the same price and beats it on all eight tests in the launch table, some by a wide margin: on Terminal-Bench 4.0, 70.6% against 10.3%; on GDPval-AA v2.1, 1,844 against 1,449. Anthropic also says 5.5 generates output more than 30% faster and costs up to 30% less per task.
On CursorBench 4.0, Sonnet 5.5 at low effort scores 35.8% for US$ 0.50 per task, above Sonnet 5's best, 34.1% for US$ 7.17 at max. With the same price per token, the savings come from spending fewer tokens.
What keeps Sonnet 5 around is code, with five changes that return errors, and its new role as Sonnet 5.5's fallback model for cybersecurity refusals. It stays available until at least Jun 30, 2027 on Anthropic's platforms; Amazon Bedrock and Google Cloud set their own dates.
| Sonnet 5 legacy | Sonnet 5.5 current | Opus 5.5 current | |
|---|---|---|---|
| Terminal-Bench 4.0 long terminal work |
10.3% | 70.6% | 66.4% |
| FrontierCode 1.1 coding |
42.4% | 52.1% | 54.4% |
| CursorBench 4.0 coding |
34.1% | 55.5% | 57.8% |
| GDPval-AA v2.1 office work, Elo |
1,449 | 1,844 | 1,846 |
| AA-Briefcase v1.1 long office work, Elo |
1,359 | 1,811 | 1,822 |
| Humanity's Last Exam with tools |
54.9% | 64.5% | 67.7% |
| OSWorld 2.1 computer use, partial |
57.0% | 80.1% | 81.8% |
| Chartography chart reading, without tools |
15.6% | 61.6% | 64.4% |
Figures from the Claude Sonnet 5.5 announcement. 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.
What breaks when you migrate to Sonnet 5.5
Swapping the identifier is not enough. Anthropic's release notes list five ways code written for Sonnet 5 breaks on 5.5:
- Thinking turned off:
disabledreturns a 400 error. The minimum is nowbetween_tools, accepted up to high effort. - Forced tool use:
tool_choicewithanyortoolreturns a 400 error. Useautowith the tool in strict mode. - Edited history: on accounts created on or after Aug 31, 2026, replaying a thinking block after editing the conversation history returns a 400 error.
- Computer use: on the Anthropic API and Google Cloud, only
computer_toolset_20260801;computer_20251124returns an error. - Advisor tool: Sonnet 5 is no longer accepted as the advisor for a Sonnet 5.5 executor.
- No error, but an effect: text between tool calls arrives in thinking blocks, empty by default, and effort was recalibrated.
Examples of using Claude Sonnet 5
A security tool runs on Sonnet 5.5 with fallbacks in default mode. When the classifier blocks a cybersecurity request, the API reruns it on Sonnet 5, and the response names the model that served it. That is why Sonnet 5 can show up on the bill even after migrating.
A pipeline reads invoices and returns fields through forced tool use, with thinking turned off to stay fast. On Sonnet 5, it works. On Sonnet 5.5, replace disabled with between_tools and forced tool use with auto in strict mode, and say in the prompt when to call the tool.
Run ten real tasks on Sonnet 5 and Sonnet 5.5, both at the API default effort, high, and compare accuracy and cost per task. With the same price per token, any difference on the bill comes only from the tokens spent.
Where Claude Sonnet 5 stands out and where it falls behind, by the numbers
Claude Sonnet 5 has published results on 26 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Oct 3, 2025 and Sep 28, 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 26 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| GPQA Diamond | 90.5% | #34 / 77 | #36 / 173 | 96.4% Gemini 4 Argon |
| AIME 2024 and 2025 (OTIS) | 94.7% | #35 / 78 | #41 / 168 | 100.0% GPT-5.5 |
| Dtbench | 92.5% | #27 / 70 | #27 / 148 | 98.9% Claude Opus 5.5 |
| Lmca | 50.02 | #21 / 70 | #21 / 121 | 68.23 Claude Opus 5.5 |
| CritPt | 16.9% | #30 / 75 | #34 / 112 | 32.3% GPT-5.6 Sol |
| WeirdML | 68.8% | #20 / 56 | #20 / 112 | 93.6% GPT-6 Astra |
| SciCode | 54.3% | #27 / 72 | #31 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 65.6% | #29 / 63 | #34 / 86 | 93.7% GPT-6 Astra |
| ALE-Bench | 1,463 | #15 / 63 | #16 / 82 | 2,951 GPT-6 Astra |
| WebDev Arena | 1,537 | #31 / 74 | #33 / 82 | 1,820 Claude Opus 5.5 |
| SimpleQA Verified | 33.7% | #48 / 60 | #56 / 81 | 76.0% Gemini 4 Argon |
| Chess puzzles | 45.0% | #23 / 62 (tied with 2) | #26 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 35.0% | #17 / 57 (tied with 1) | #20 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 77.0% | #10 / 65 | #14 / 74 | 100.0% Claude Fable 5.1 |
| SimpleBench | 60.6% | #23 / 39 | #24 / 74 | 81.9% Claude Fable 5 |
| Frontiermath Tier 4 V2 | 29.3% | #26 / 53 (tied with 2) | #31 / 67 | 100.0% GPT-6.1 Sol |
| Vending Bench 2 | 6,378 | #16 / 47 | #18 / 57 | 15,515 GPT-6 Astra |
| Apex Agents | 54.5% | #18 / 46 | #21 / 50 | 82.4% Gemini 4 Argon |
| FrontierCode | 42.7% | #17 / 39 | #20 / 43 | 54.6% Claude Opus 5.5 |
| Blueprint Bench 2 | 24.9% | #20 / 29 | #22 / 31 | 54.4% Gemini 4 Argon |
| Gso | 37.3% | #9 / 17 | #9 / 28 | 88.2% Claude Fable 5.1 |
| DeepSWE | 53.8% | #17 / 26 | #17 / 26 | 74.1% GPT-6 Astra |
| Surface Evolver Bench | 60.0% | #10 / 22 | #10 / 24 | 95.0% Claude Fable 5 |
| Gbaeval | 65.3% | #5 / 22 | #5 / 22 | 79.6% Claude Opus 5 |
| CursorBench | 34.1% | #14 / 14 | #16 / 16 | 57.8% Claude Opus 5.5 |
| Btf3 | 14.2% | #4 / 8 | #4 / 8 | 14.9% GLM-5.3 |
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 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 |
|---|---|---|---|---|---|
| Qwen 3.8 Max | Alibaba | 156.88 | 1,536 | 78.5 | Aug 2, 2026 |
| Muse Spark 1.3 | Meta AI | 156.76 | 1,528 | 81.6 | Sep 2, 2026 |
| Claude Sonnet 5 | Anthropic | 156.36 | 1,491 | 76.0 | Jun 30, 2026 |
| GPT-5.6 Luna | OpenAI | 156.33 | 1,478 | 73.6 | Jul 9, 2026 |
| Grok 4.6 | xAI | 156.31 | 1,475 | 78.0 | Aug 12, 2026 |
| GPT-6 Luna | OpenAI | 156.31 | n/a | 72.0 | Sep 22, 2026 |
| GLM-5.3 | Z.ai (Zhipu AI) | 155.91 | n/a | 76.1 | Aug 14, 2026 |
Claude Sonnet 5 or GPT-5.6 Luna?
GPT-5.6 Luna is the model from another lab closest to Claude Sonnet 5 in capability. Across the 22 tests both took, this is the score:
| Test | Claude Sonnet 5 | GPT-5.6 Luna |
|---|---|---|
| Frontiermath Tier 4 V2 | 29.3% | 61.0% |
| Game puzzles | 35.0% | 21.0% |
| Vending Bench 2 | 6,378 | 4,095 |
| SimpleBench | 60.6% | 46.8% |
| ProofBench | 77.0% | 60.0% |
| Apex Agents | 54.5% | 43.0% |
| Frontiermath Tiers 1 3 V2 | 65.6% | 82.1% |
| DeepSWE | 53.8% | 67.2% |
| CritPt | 16.9% | 20.6% |
| SimpleQA Verified | 33.7% | 41.0% |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
What Claude Sonnet 5 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.
DeepSWE
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? 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
Does Claude Sonnet 5 still work?
Yes. It is a legacy model, but it is still available on the Anthropic API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. Anthropic guarantees it will not be retired before June 30, 2027 on the platforms it operates, with at least 60 days' notice.
How much does Claude Sonnet 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. It is exactly the Sonnet 5.5 price list.
Will Sonnet 5 go up to US$ 3 and US$ 15?
No. The increase was scheduled for September 1, 2026, but on August 10, 2026 Anthropic made the launch price of US$ 2 and US$ 10 permanent. The cost charts in the announcement still use the old prices.
Sonnet 5 or Sonnet 5.5?
For new projects, Sonnet 5.5: it costs the same, wins all eight tests in the launch table and, according to Anthropic, generates output more than 30% faster. Sonnet 5 only makes sense while your code depends on features 5.5 rejects, such as turning thinking off with disabled and forced tool use.
Why was a request to Sonnet 5.5 answered by Sonnet 5?
Because Sonnet 5.5's safety classifier refused the request, usually a cybersecurity one, and it fell to the fallback model. On the API, this happens with the fallbacks parameter in default mode, and that part is billed at Sonnet 5 rates.
Where does Claude Sonnet 5 rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, Claude Sonnet 5 ranks #24 out of 267 models, with an ECI of 156.36. 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 better than GPT-5.6 Luna?
Across the 22 tests both models took and Epoch AI compiled, Claude Sonnet 5 comes out ahead on 12, behind on 10 and tied on 0. 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 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 are: ProofBench (#10 of 65); Gbaeval (#5 of 22); ALE-Bench (#15 of 63). 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 announcement
- Anthropic, Claude Sonnet 5 system card
- Anthropic, Claude Sonnet 5 overview (documentation)
- Anthropic, prompting Claude Sonnet 5 (documentation)
- Anthropic, pricing (documentation)
- Anthropic, model deprecations (documentation)
- Anthropic, migrating to Claude Sonnet 5.5 (documentation)
- Anthropic, refusals and fallback (documentation)
- Anthropic, platform release notes (documentation)
- Anthropic, Claude Sonnet 5.5 announcement (comparison with Sonnet 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.
← See the full ranking