Up to four models
Gemini 3.7 Flash vs Gemini 3.6 Flash
On the index (GPT-4 = 100), Gemini 3.7 Flash scores 170 and Gemini 3.6 Flash 161. Side by side on the same third-party tests, with price and context. The link in your address bar shares this exact comparison.
170154–186
WeightsClosed
ReleasedAug 2026
Price in / out$0.75 / $3.75
Context1.05M
Core tests taken6 of 9
161145–176
WeightsClosed
ReleasedJul 2026
Price in / out$0.75 / $3.75
Context1.05M
Core tests taken5 of 9
Gemini 3.7 Flash leads Gemini 3.6 Flash by 10 points, ahead on 5 of 5 shared core tests, at a similar price.
Test by testthird-party results
Every result as a table
| Test | Gemini 3.7 Flash | Gemini 3.6 Flash |
|---|---|---|
| Reasoning & knowledge | ||
| GPQA Diamond (Epoch AI) | 94.8% | 94.1% |
| GPQA Diamond (Vals AI) | 93.9% | 93.4% |
| MMLU Pro | 90.1% | 89.3% |
| MMMU Pro | 89.0% | 88.4% |
| SimpleQA Verified | 69.2% | 66.2% |
| Maths | ||
| FrontierMath (tiers 1-3) | 71.6% | 59.0% |
| AIME-style maths (OTIS mock) | 97.2% | 94.2% |
| Chess puzzles | 66.0% | 42.0% |
| FrontierMath (tier 4) | 36.6% | 21.9% |
| Coding | ||
| Code Migration | 34.8% | 30.9% |
| CyberBench v1.1 | 43.8% | 45.4% |
| IOI | 67.8% | 35.1% |
| LiveCodeBench | 88.6% | 88.1% |
| ProgramBench | 0.0% | 0.0% |
| SWE-bench | 80.8% | 79.6% |
| Vibe Code Bench v1.1 | 70.4% | 64.0% |
| DeepSWE | 65.5% | 46.7% |
| MirrorCode | 41.1% | 23.3% |
| Agents | ||
| APEX-Agents | 67.8% | 46.9% |
| Terminal-Bench-Science | 7.1% | 4.3% |
| DTBench | 96.8% | 95.5% |
| Furniture assembly | 26.7% | 23.3% |
| LMCA | 50.4% | 44.9% |
| Mystery game puzzles | 37.0% | 30.0% |
| Professional work | ||
| Finance Agent (v2) | 59.0% | 56.3% |
| EMB | 71.3% | 65.4% |
| LegalBench | 87.3% | 86.7% |
| Legal Research Bench | 34.6% | 25.0% |
| MedCode | 53.4% | 53.1% |
| MedScribe | 83.9% | 79.7% |
| MortgageTax | 66.6% | 67.6% |
| Tax Agent Bench | 22.8% | 18.0% |
| TaxEval v2 | 74.7% | 74.9% |
| Novel problems | ||
| ARC-AGI-2 | 84.6% | 60.4% |
| ARC-AGI-1 | 95.5% | 91.2% |
Tests at least two of these models have taken. Bold rows are the core tests.