Gemini 3 Flash (Preview) (High): benchmark results
Gemini 3 Flash (Preview) evaluated at the high reasoning-effort setting. Provider: Google. Released 2025-12-17. Access: API.
Aggregate Rating 1755 ± 11, rank #226 of 2168 rated models, from 87 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Context Arena MRCR (2-needle) | 93.8 | AUC@1M (%) | 100 |
| Context Arena MRCR (4-needle) | 68 | AUC@1M (%) | 100 |
| Context Arena MRCR (8-needle) | 49.4 | AUC@1M (%) | 100 |
| DocVQA 2026 - Comics | 40 | Accuracy (%, high thinking) | 100 |
| DocVQA 2026 - Engineering Drawing | 50 | Accuracy (%, high thinking) | 100 |
| DocVQA 2026 - Science Paper | 50 | Accuracy (%, high thinking) | 100 |
| ImmersedPrivacy - Social Context Selection (Native Audio) | 54 | Selection accuracy (%; Tier 2: share of trials in which the | 100 |
| Vals AI MedCode | 55.92 | Accuracy (%) | 96.3 |
| UGI - Natural Intelligence | 68.85 | NatInt Score | 95.7 |
| CiteVQA - Evidence Recall | 45.4 | Recall (%) of the crucial evidence elements matched by a cit | 94.7 |
| CiteVQA - Evidence Relevance | 75.7 | Relevance of each cited region to its answer, judged on a 0- | 94.7 |
| CiteVQA - Multi-Doc (1-Gold) | 61.8 | Strict attributed accuracy (%): share of questions with answ | 94.7 |
Interactive version: theaggregate.ai/model?slug=gemini-3-flash-preview-high · How It Works · Data refreshed daily, snapshot 2026-10-20.