Gemini 3 Flash (Preview): benchmark results
Google preview Gemini 3 Flash model, a speed-oriented Flash-tier Gemini 3 snapshot. Provider: Google. Released 2025-12-17. Access: API.
Aggregate Rating 1779 ± 3, rank #82 of 1656 rated models, from 1757 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| ATC Safety-Oriented Language Understanding Eval | 69.22 | Risk Score (imported) | 100 |
| AgentProcessBench | 81.6 | StepAcc (%): share of assistant steps, pooled over trajector | 100 |
| AgentProcessBench - BFCL First-Error Accuracy | 64 | FirstErrAcc (%): share of trajectories where the judge locat | 100 |
| AgentProcessBench - BFCL Step Accuracy | 81.8 | StepAcc (%): share of assistant steps, pooled over trajector | 100 |
| AgentProcessBench - First-Error Accuracy | 65.8 | FirstErrAcc (%): share of trajectories where the judge locat | 100 |
| AgentProcessBench - GAIA First-Error Accuracy | 65.2 | FirstErrAcc (%): share of trajectories where the judge locat | 100 |
| AgentProcessBench - GAIA Step Accuracy | 79.7 | StepAcc (%): share of assistant steps, pooled over trajector | 100 |
| AgentProcessBench - HotPotQA Step Accuracy | 76.7 | StepAcc (%): share of assistant steps, pooled over trajector | 100 |
| AgentProcessBench - tau2-Bench Step Accuracy | 83.4 | StepAcc (%): share of assistant steps, pooled over trajector | 100 |
| AgentWebBench (Centralized Retrieval) - Question Answering | 60.38 | Answer accuracy (%, LLM judge) on 53 DeepResearchGym short-a | 100 |
| AgentWebBench (Centralized Retrieval) - Web Search | 53.77 | NDCG@5 (%) on 354 MS MARCO Web Search queries whose relevant | 100 |
| AgentWebBench (Embedding-Selected Tools) - Question Answering | 56.6 | Answer accuracy (%, LLM judge) on 53 DeepResearchGym short-a | 100 |
Interactive version: theaggregate.ai/model?slug=gemini-3-flash-preview · How It Works · Data refreshed daily, snapshot 2026-10-20.