Across 12 shared benchmarks, Claude Sonnet 4.6 leads overall: Claude Sonnet 4.6 wins 9, Claude Sonnet 4.5 wins 3, with 0 ties and an average score difference of +18.09.
Anthropic · 2026-02-17 · AI model
Anthropic · 2025-09-30 · AI model
Grouped by capability, sorted by largest gap within each. 12 shared benchmarks.
| Benchmark | Claude Sonnet 4.6 | Claude Sonnet 4.5 | Diff |
|---|---|---|---|
| ARC-AGI-2 | 58.3017 / 58thinking | 3.8048 / 58 | +54.50 |
| HLE | 4920 / 149thinking + 使用工具 | 7.10136 / 149 | +41.90 |
| GPQA Diamond | 89.9018 / 175thinking | 73.7094 / 175 | +16.20 |
| Benchmark | Claude Sonnet 4.6 | Claude Sonnet 4.5 | Diff |
|---|---|---|---|
| Terminal Bench 2.0 | 59.1019 / 43thinking + 使用工具 | 42.8038 / 43thinking + 使用工具 | +16.30 |
| OSWorld-Verified | 72.507 / 14thinking + 使用工具 | 61.4010 / 14thinking + 使用工具 | +11.10 |
| Benchmark | Claude Sonnet 4.6 | Claude Sonnet 4.5 | Diff |
|---|---|---|---|
| τ²-Bench - Telecom | 97.909 / 35thinking + 使用工具 | 985 / 35thinking + 使用工具 | -0.10 |
| Benchmark | Claude Sonnet 4.6 | Claude Sonnet 4.5 | Diff |
|---|---|---|---|
| BrowseComp | 74.7018 / 43thinking + 使用工具 | 24.1041 / 43thinking + 使用工具 | +50.60 |
| Benchmark | Claude Sonnet 4.6 | Claude Sonnet 4.5 | Diff |
|---|---|---|---|
| Pinch Bench | 885 / 37Thinking (With Tools) | 88.204 / 37Thinking (With Tools) | -0.20 |
| Benchmark | Claude Sonnet 4.6 | Claude Sonnet 4.5 | Diff |
|---|---|---|---|
| SWE-bench Verified | 79.6013 / 103thinking | 823 / 103parallel_thinking + 使用工具 | -2.40 |
| Benchmark | Claude Sonnet 4.6 | Claude Sonnet 4.5 | Diff |
|---|---|---|---|
| AA-LCR | 711 / 13thinking | 668 / 13thinking | +5 |
| Benchmark | Claude Sonnet 4.6 | Claude Sonnet 4.5 | Diff |
|---|---|---|---|
| FrontierMath - Tier 4 | 8.3034 / 80Thinking (No Tools, 16K Budget) | 2.1056 / 80Normal (No Tools) | +6.20 |
| Benchmark | Claude Sonnet 4.6 | Claude Sonnet 4.5 | Diff |
|---|---|---|---|
| GDPval-AA | 5710 / 20thinking | 3915 / 20thinking | +18 |
| Field | Claude Sonnet 4.6 | Claude Sonnet 4.5 |
|---|---|---|
| Publisher | Anthropic | Anthropic |
| Release date | 2026-02-17 | 2025-09-30 |
| Model type | AI model | AI model |
| Architecture | Dense | Dense |
| Parameters | 0.0 | 0.0 |
| Context length | 1M | 1000K |
| Max output | 8192 | 65536 |
Prices use DataLearner records when available; missing fields are not inferred.
| Item | Claude Sonnet 4.6 | Claude Sonnet 4.5 |
|---|---|---|
| Text input | $3 / 1M tokens | 3 美元/100 万tokens |
| Text output | $15 / 1M tokens | 15 美元/100 万tokens |
| Cache read | $0.3 / 1M tokens | 3.75 美元/100 万tokens |
| Cache write | $3.75 / 1M tokens | 0.3 美元/100 万tokens |
On average across the 12 shared benchmarks, Claude Sonnet 4.6 scores 18.09 higher.
Largest single-benchmark gap: ARC-AGI-2 — Claude Sonnet 4.6 58.30 vs Claude Sonnet 4.5 3.80 (+54.50).
Page generated from structured model, pricing and benchmark records. No real-time LLM is used to write the prose.
