AnthropicReasoning modelClaude Sonnet 4
Best configuration ranks #208 of 610 on the BenchLeader Index at 54.0 ±3.1 (thinking reasoning effort). Last measured 2 Sept 2026. Released 22 May 2025.
- Blended price
- $6.00/M
- $3.00 in · $15.00 out
- Output speed
- 20 tok/s
- First answer
- 1.14 s
- Context
- 1M
- Overall index54
- Reasoning48
- Coding46
- Agents & tools61
- Maths55
- Knowledge58
- Instruction following53
- Human preference58
- Multimodal51
- Long context62
- Composite53
Reasoning-effort configurations
The same model behaves differently depending on how much it is allowed to think. Each row is one setting, scored only on the benchmarks that were run at that setting. “Default” means the publisher did not say which setting was used.
| Effort | Index | Rank | Speed | First answer | Chat reply cost | Categories |
|---|---|---|---|---|---|---|
| thinkingbest | 54.0 | #208 | 20 tok/s | 1.14 s | $0.0057 | Agents & tools 61 · Coding 46 · Composite 53 · Human preference 58 · Instruction following 53 · Knowledge 58 · Long context 62 · Maths 55 · Multimodal 51 · Reasoning 48 |
| default | 49.5 | #305 | 20 tok/s | 1.14 s | $0.0057 | Agents & tools 48 · Coding 51 · Composite 50 · Human preference 57 · Instruction following 47 · Knowledge 51 · Long context 42 · Maths 54 · Multimodal 51 · Reasoning 45 |
Benchmark results
One column per reasoning effort. Rank is among every configuration of every model on that benchmark. Hover a score for the run it came from.
Reasoning
| Benchmark | thinking | default | Source |
|---|---|---|---|
| GPQA Diamond | – | 79.2%#117 | Epoch AI Benchmarking Hub |
| Humanity's Last Exam | – | 7.8%#33 | Scale AI / CAIS |
| SimpleBench | – | 45.5%#55 | SimpleBench |
| LMArena Hard Prompts | 1432#116 | 1419#132 | LMArena |
| GPQA Diamond (AA)not in index | 77.7%#211 | 68.3%#300 | Artificial Analysis |
| Humanity's Last Exam (AA)not in index | 10.7%#247 | 4.3%#434 | Artificial Analysis |
| GPQA Diamond (Vals)not in index | 75.0%#79 | 69.4%#92 | Vals AI |
| Kagi LLM Benchmark | 73.0%#20 | 55.9%#61 | Kagi LLM Benchmark |
| ARC-AGI-1 | 40.0%#132 | 23.8%#151 | ARC Prize |
| ARC-AGI-2 | 5.9%#116 | 1.3%#153 | ARC Prize |
Coding
| Benchmark | thinking | default | Source |
|---|---|---|---|
| SciCode | – | 40.0%#106 | SciCode |
| WeirdML | – | 46.1%#77 | WeirdML |
| GSO-Bench | – | 4.9%#21 | GSO-Bench |
| LMArena Coding | 1474#98 | 1450#127 | LMArena |
| LiveCodeBench | 62.4%#100 | 59.7%#103 | Vals AI |
| IOI | 4.6%#45 | 6.5%#40 | Vals AI |
| SWE-Bench Pro | – | 42.7%#9 | Scale AI SEAL |
| Aider Polyglot | – | 61.3%#13 | Aider polyglot leaderboard |
| SWE-bench Verified (bash only) | – | 64.9%#20 | SWE-bench |
| SWE-bench Verified (any scaffold)not in index | – | 76.8%#4 | SWE-bench |
Agents & tools
| Benchmark | thinking | default | Source |
|---|---|---|---|
| Cybench | – | 35.0%#8 | Cybench |
| APEX-Agents | – | 9.3%#52 | Mercor |
| Terminal-Bench Hard | 31.1%#104 | 27.3%#123 | Artificial Analysis |
| τ²-Bench Telecom (AA)not in index | 64.6%#166 | 52.3%#188 | Artificial Analysis |
Maths
| Benchmark | thinking | default | Source |
|---|---|---|---|
| OTIS Mock AIME | – | 71.1%#120 | Epoch AI Benchmarking Hub |
| MATH Level 5 | – | 84.4%#27 | Epoch AI Benchmarking Hub |
| AIME (Vals) | 76.3%#51 | 38.5%#69 | Vals AI |
| MGSM | 90.9%#32 | 93.0%#11 | Vals AI |
Knowledge
| Benchmark | thinking | default | Source |
|---|---|---|---|
| AA-Omniscience | 0.2#94 | -9.0#133 | Artificial Analysis |
| MMLU-Pro | 83.9%#67 | 79.4%#91 | Vals AI |
| LegalBench | 82.1%#68 | 83.0%#56 | Vals AI |
| CorpFin | 61.2%#55 | 54.7%#89 | Vals AI |
| TaxEval | 72.0%#68 | 69.6%#94 | Vals AI |
| MedQA | 92.7%#27 | 90.3%#45 | Vals AI |
| MultiNRC | – | 18.4%#34 | Scale AI SEAL |
Instruction following
| Benchmark | thinking | default | Source |
|---|---|---|---|
| IFBench | 54.7%#147 | 45.4%#202 | Artificial Analysis |
| MultiChallenge | 57.1%#15 | – | Scale AI SEAL |
Human preference
| Benchmark | thinking | default | Source |
|---|---|---|---|
| LMArena Text | 1401#131 | 1390#142 | LMArena |
Multimodal
| Benchmark | thinking | default | Source |
|---|---|---|---|
| LMArena Vision | 1191#79 | 1175#87 | LMArena |
| MMMU-Pro | 61.8%#171 | 62.4%#162 | Artificial Analysis |
| VISTA | 45.5%#24 | 43.2%#34 | Scale AI SEAL |
| MMMU (validation) | – | 74.4%#16 | MMMU |
Long context
| Benchmark | thinking | default | Source |
|---|---|---|---|
| Fiction.LiveBench 120k | – | 36.4%#29 | Fiction.live |
| AA-LCR | 70.3%#157 | 44.0%#272 | Artificial Analysis |
Composite
| Benchmark | thinking | default | Source |
|---|---|---|---|
| Epoch Capabilities Indexnot in index | – | 141.7#91 | Epoch AI Benchmarking Hub |
| AA Intelligence Index | 18.9#185 | 16.6#209 | Artificial Analysis |
Where to run it
Every provider serving this model through OpenRouter, with throughput and first-token latency measured on live traffic over the last 30 minutes and each provider’s own price. Purple marks the best in each column.
| Provider | Speed | First token | Input $/M | Output $/M | Context | Quantisation |
|---|---|---|---|---|---|---|
| Amazon Bedrock | 35 tok/s | 1.14 s | $3.00 | $15.00 | 200k | – |
| Google Vertex (Global) | 20 tok/s | 1.11 s | $3.00 | $15.00 | 1M | – |
| Amazon Bedrock (EU) | 3 tok/s | 1.70 s | $3.00 | $15.00 | 200k | – |
What a task costs
Estimates from list price, output speed and time to first answer for the best configuration. “With caching” assumes three-quarters of the input is served from the prompt cache at $0.300 per 1M. Reasoning tokens are not modelled.
| Workload | Tokens in / out | Cost | With caching | Time |
|---|---|---|---|---|
| Chat reply | 400 / 300 | $0.0057 | $0.0049 | 16.1 s |
| Summarise a 30-page report | 12,000 / 600 | $0.045 | $0.021 | 31.1 s |
| Code edit | 6,000 / 1,500 | $0.041 | $0.028 | 1.3 min |
| Agentic coding session | 60,000 / 4,000 | $0.240 | $0.118 | 3.4 min |
| Structured extraction | 2,000 / 200 | $0.0090 | $0.0050 | 11.1 s |
See also
Data as of 9 Sept 2026. Compare these configurations.