Claude Sonnet 4.6vsClaude Opus 4.6
Across 14 shared benchmarks, Claude Opus 4.6 leads overall: Claude Sonnet 4.6 wins 5, Claude Opus 4.6 wins 9, with 0 ties and an average score difference of -1.79.
Claude Sonnet 4.6
Anthropic · 2026-02-17 · Chat model
Claude Opus 4.6
Anthropic · 2026-02-05 · Reasoning model
Claude Sonnet 4.65 wins(36%)(64%)9 winsClaude Opus 4.6
Benchmark scores
Grouped by capability, sorted by largest gap within each. 14 shared benchmarks.
Agent Level Benchmark
Claude Opus 4.6 2/3| Benchmark | Claude Sonnet 4.6 | Claude Opus 4.6 | Diff |
|---|---|---|---|
| τ³-Banking | 34.4049 / 164Max (With Tools) | 27.3269 / 164Max (With Tools) | +7.08 |
| τ²-Bench - Telecom | 79.50115 / 264Normal (With Tools) | 84.8090 / 264Normal (With Tools) | -5.30 |
| Terminal Bench Hard | 46.2029 / 244Normal (With Tools) | 48.5025 / 244Normal (With Tools) | -2.30 |
General Knowledge
Claude Opus 4.6 2/3| Benchmark | Claude Sonnet 4.6 | Claude Opus 4.6 | Diff |
|---|---|---|---|
| HLE | 13.30355 / 563Normal (No Tools) · Text only | 19.10299 / 563Normal (No Tools) · Text only | -5.80 |
| CritPt | 0.90158 / 200Normal (No Tools) | 2.80120 / 200Normal (No Tools) | -1.90 |
| LiveBench | 75.3211 / 117Thinking High (No Tools) | 74.5217 / 117Thinking High (No Tools) | +0.80 |
Claw-style Agent Evaluation
Even 2/2| Benchmark | Claude Sonnet 4.6 | Claude Opus 4.6 | Diff |
|---|---|---|---|
| PinchBench v2 | 62.6831 / 45Reported best (effort unspecified) | 69.9026 / 45Reported best (effort unspecified) | -7.22 |
| Pinch Bench | 886 / 38Thinking (With Tools) | 87.408 / 38Thinking (With Tools) | +0.60 |
General Evaluation
Claude Opus 4.6 1/1| Benchmark | Claude Sonnet 4.6 | Claude Opus 4.6 | Diff |
|---|---|---|---|
| GPQA Diamond | 79.90237 / 462Normal (No Tools) | 84176 / 462Normal (No Tools) | -4.10 |
Instruction Following
Claude Opus 4.6 1/1| Benchmark | Claude Sonnet 4.6 | Claude Opus 4.6 | Diff |
|---|---|---|---|
| IF Bench | 42.40201 / 282Normal (No Tools) | 44.60182 / 282Normal (No Tools) | -2.20 |
Long Context
Claude Sonnet 4.6 1/1| Benchmark | Claude Sonnet 4.6 | Claude Opus 4.6 | Diff |
|---|---|---|---|
| AA-LCR | 69.30107 / 170Normal (No Tools) | 67116 / 170Normal (No Tools) | +2.30 |
Multimodal Understanding
Claude Opus 4.6 1/1| Benchmark | Claude Sonnet 4.6 | Claude Opus 4.6 | Diff |
|---|---|---|---|
| MMMU-Pro | 70.60123 / 227Normal (No Tools) | 72.50114 / 227Normal (No Tools) | -1.90 |
Text Embedding
Claude Opus 4.6 1/1| Benchmark | Claude Sonnet 4.6 | Claude Opus 4.6 | Diff |
|---|---|---|---|
| Context Arena | 55.1880 / 126Normal (No Tools) | 60.3376 / 126Normal (No Tools) | -5.15 |
Writing and Creative Capabilities
Claude Sonnet 4.6 1/1| Benchmark | Claude Sonnet 4.6 | Claude Opus 4.6 | Diff |
|---|---|---|---|
| Creative Writing | 1,80418 / 106Normal (No Tools) | 1,80419 / 106Normal (No Tools) | +0.10 |
Specs
| Field | Claude Sonnet 4.6 | Claude Opus 4.6 |
|---|---|---|
| Publisher | Anthropic | Anthropic |
| Release date | 2026-02-17 | 2026-02-05 |
| Model type | Chat model | Reasoning model |
| Architecture | Dense | Dense |
| Parameters | Not available | Not available |
| Context length | 1M | 1000K |
| Max output | 8K | 64K |
API pricing
Prices use DataLearner records when available; missing fields are not inferred.
| Item | Claude Sonnet 4.6 | Claude Opus 4.6 |
|---|---|---|
| Text input | $3 / 1M tokens | $5 / 1M tokens |
| Text output | $15 / 1M tokens | $25 / 1M tokens |
| Cache read | $0.3 / 1M tokens | $0.5 / 1M tokens |
| Cache write | $3.75 / 1M tokens | $6.25 / 1M tokens |
Summary
- Claude Sonnet 4.6leads in:Long Context (1/1), Writing and Creative Capabilities (1/1)
- Claude Opus 4.6leads in:Agent Level Benchmark (2/3), General Knowledge (2/3), General Evaluation (1/1), Instruction Following (1/1), Multimodal Understanding (1/1), Text Embedding (1/1)
- Tied in:Claw-style Agent Evaluation
On average across the 14 shared benchmarks, Claude Opus 4.6 scores 1.79 higher.
Largest single-benchmark gap: PinchBench v2 — Claude Sonnet 4.6 62.68 vs Claude Opus 4.6 69.90 (-7.22).
Page generated from structured model, pricing and benchmark records. No real-time LLM is used to write the prose.