Up to four models
Claude Sonnet 5 vs DeepSeek-V4-Pro
On the index (GPT-4 = 100), Claude Sonnet 5 scores 164 and DeepSeek-V4-Pro 163. Side by side on the same third-party tests, with price and context. The link in your address bar shares this exact comparison.
Claude Sonnet 5
164148–180
WeightsClosed
ReleasedJun 2026
Price in / out$2.0 / $10
Context1M
Core tests taken5 of 9
163147–178
WeightsOpen
ReleasedAug 2026
Price in / out$0.11 / $4.44
Context1.05M
Core tests taken6 of 9
Claude Sonnet 5 and DeepSeek-V4-Pro are too close to call (2 points apart, within the uncertainty), ahead on 3 of 5 shared core tests, at 3.4× the price.
Test by testthird-party results
Claude Sonnet 5
DeepSeek-V4-Pro
Every result as a table
| Test | Claude Sonnet 5 | DeepSeek-V4-Pro |
|---|---|---|
| Reasoning & knowledge | ||
| GPQA Diamond (Epoch AI) | 90.5% | 91.7% |
| LiveBench Data Analysis | 71.7% | 79.2% |
| LiveBench IF | 63.9% | 67.7% |
| LiveBench Language | 75.0% | 82.1% |
| LiveBench Reasoning | 88.7% | 85.8% |
| GPQA Diamond (Vals AI) | 88.9% | 92.4% |
| MMLU Pro | 87.5% | 87.0% |
| SimpleQA Verified | 33.7% | 52.9% |
| Maths | ||
| FrontierMath (tiers 1-3) | 65.6% | 64.6% |
| AIME-style maths (OTIS mock) | 94.7% | 98.6% |
| LiveBench Mathematics | 92.9% | 95.1% |
| ProofBench v1.1 | 77.0% | 50.0% |
| Chess puzzles | 45.0% | 46.0% |
| FrontierMath (tier 4) | 29.3% | 26.8% |
| Coding | ||
| Terminal-Bench 4.0 | 10.6% | 14.1% |
| LiveBench Agentic Coding | 59.4% | 54.9% |
| LiveBench Coding | 80.7% | 77.2% |
| Code Migration | 44.4% | 41.5% |
| CyberBench v1.1 | 70.3% | 72.9% |
| IOI | 45.0% | 51.6% |
| LiveCodeBench | 82.4% | 87.5% |
| ProgramBench | 0.0% | 0.0% |
| SWE-bench | 79.6% | 96.4% |
| Vibe Code Bench 1-100 | 13.8% | 17.5% |
| Vibe Code Bench v1.1 | 81.3% | 82.3% |
| MirrorCode | 63.3% | 33.3% |
| WeirdML | 69.0% | 66.2% |
| Agents | ||
| APEX-Agents | 54.5% | 47.3% |
| SkillsBench | 46.5% | 53.8% |
| SRE Bench | 0.4% | 1.9% |
| Terminal-Bench-Science | 5.7% | 4.3% |
| DTBench | 92.5% | 93.9% |
| LMCA | 50.0% | 45.5% |
| Mystery game puzzles | 35.0% | 43.0% |
| Professional work | ||
| Finance Agent (v2) | 53.9% | 50.4% |
| CorpFin v2 | 67.0% | 65.4% |
| EMB | 66.3% | 52.8% |
| LegalBench | 83.9% | 82.4% |
| Legal Research Bench | 41.8% | 40.9% |
| MedCode | 47.5% | 42.5% |
| MedScribe | 76.0% | 80.2% |
| Tax Agent Bench | 29.4% | 26.8% |
| TaxEval v2 | 75.6% | 73.1% |
Tests at least two of these models have taken. Bold rows are the core tests.