Up to four models
DeepSeek-V4-Pro vs Gemini 3.1 Pro
On the index (GPT-4 = 100), DeepSeek-V4-Pro scores 163 and Gemini 3.1 Pro 161. Side by side on the same third-party tests, with price and context. The link in your address bar shares this exact comparison.
163147–178
WeightsOpen
ReleasedAug 2026
Price in / out$0.11 / $4.44
Context1.05M
Core tests taken6 of 9
161146–177
WeightsClosed
ReleasedFeb 2026
Price in / out$2.0 / $12
Context1.05M
Core tests taken9 of 9
DeepSeek-V4-Pro and Gemini 3.1 Pro are too close to call (1 point apart, within the uncertainty), ahead on 4 of 6 shared core tests, and costs 3.8× less.
Test by testthird-party results
Every result as a table
| Test | DeepSeek-V4-Pro | Gemini 3.1 Pro |
|---|---|---|
| Reasoning & knowledge | ||
| GPQA Diamond (Epoch AI) | 91.7% | 94.4% |
| GPQA Diamond (Vals AI) | 92.4% | 95.5% |
| MMLU Pro | 87.0% | 91.0% |
| SimpleQA Verified | 52.9% | 73.5% |
| Maths | ||
| FrontierMath (tiers 1-3) | 64.6% | 59.7% |
| AIME-style maths (OTIS mock) | 98.6% | 95.6% |
| ProofBench v1.1 | 50.0% | 26.0% |
| Chess puzzles | 46.0% | 51.0% |
| FrontierMath (tier 4) | 26.8% | 26.8% |
| Coding | ||
| Terminal-Bench 4.0 | 14.1% | 2.5% |
| Code Migration | 41.5% | 17.3% |
| IOI | 51.6% | 51.8% |
| LiveCodeBench | 87.5% | 88.5% |
| ProgramBench | 0.0% | 0.0% |
| SWE-bench | 96.4% | 78.8% |
| Vibe Code Bench 1-100 | 17.5% | 6.7% |
| Vibe Code Bench v1.1 | 82.3% | 32.0% |
| MirrorCode | 33.3% | 8.9% |
| WeirdML | 66.2% | 72.1% |
| Agents | ||
| APEX-Agents | 47.3% | 35.1% |
| SRE Bench | 1.9% | 0.0% |
| Terminal-Bench-Science | 4.3% | 1.4% |
| DTBench | 93.9% | 97.1% |
| LMCA | 45.5% | 53.8% |
| Mystery game puzzles | 43.0% | 34.0% |
| Professional work | ||
| Finance Agent (v2) | 50.4% | 43.0% |
| CorpFin v2 | 65.4% | 64.5% |
| EMB | 52.8% | 52.6% |
| LegalBench | 82.4% | 87.4% |
| Legal Research Bench | 40.9% | 20.7% |
| MedCode | 42.5% | 59.1% |
| MedScribe | 80.2% | 76.1% |
| Tax Agent Bench | 26.8% | 9.3% |
| TaxEval v2 | 73.1% | 72.9% |
| Novel problems | ||
| ARC-AGI-2 | 61.3% | 77.1% |
| ARC-AGI-1 | 90.5% | 98.0% |
Task length · METR
46 sec
Tests at least two of these models have taken. Bold rows are the core tests.