Gemini 3.1 Flash: benchmark results
Google's Gemini 3.1 Flash, the fast mid-tier of the Gemini 3.1 generation (March 2026). Provider: Google. Released 2026-03-03. Access: API.
Aggregate Rating 1710 ± 18, rank #300 of 10916 rated models, from 75 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| NextMotionQA | 58.44 | Overall score (0-100): macro-average of the three task score | 100 |
| NextMotionQA - Captioning | 43.53 | Task 2 captioning score (0-100): Qwen3.6-Plus judge rating o | 100 |
| NextMotionQA - Error Correction | 68.23 | Task 3 caption error correction score (0-100): mean of error | 100 |
| RoboProcessBench - Current Primitive Recognition | 38.4 | Accuracy (%) on 359 questions asking which low-level primiti | 100 |
| RoboProcessBench - Next Primitive Prediction | 67.5 | Accuracy (%) on 200 questions asking which low-level primiti | 100 |
| SMH-Bench (Interactive Agent) - State Query | 95 | Success rate (%) on Environment-Grounded Query (TC7, 100 non | 100 |
| SpatialUAV - Region Recognition | 72.1 | Region Recognition Score (%; region-ID sets, partial credit, | 100 |
| HarmVideoBench | 81.3 | Macro Avg. (self-reported) | 94.1 |
| HarmVideoBench - Beyond-Clip Reasoning | 64.4 | Beyond-Clip Reasoning accuracy (%) on four-option multiple-c | 94.1 |
| HarmVideoBench - Clip-Internal Meaning | 83.7 | Clip-Internal Meaning accuracy (%) on four-option multiple-c | 94.1 |
| HarmVideoBench - Observable Evidence | 95.7 | Observable Evidence accuracy (%) on four-option multiple-cho | 94.1 |
| SpatialUAV - Aerial-Aerial Shared Association | 31.6 | Shared Association Pair F1 (%; region-pair correspondences b | 94.1 |
Interactive version: theaggregate.ai/model?slug=gemini-3-1-flash · How It Works · Data refreshed daily, snapshot 2026-11-08.