Rankings / Google DeepMind
Gemini 3 Flash Preview
released 2025-12-17
BenchAtlas Index
as of 2026-07-14
thinking
rank #50
10 families · 5 categories · high
high effort
rank #70
10 families · 4 categories · medium
base configuration
rank #115
5 families · 3 categories · medium
Benchmark evidence
71 results
agentic coding
| τ²-Bench | thinking | 80.4 % | independent | Artificial Analysis |
| τ²-Bench | no reasoning | 43.3 % | independent | Artificial Analysis |
| τ²-Bench subset=banking | thinking | 17.5 % | independent | Artificial Analysis |
coding
| LiveCodeBench v6 implementation=artificial-analysis | thinking | 90.8 % | independent | Artificial Analysis |
| LiveCodeBench v6 implementation=artificial-analysis | no reasoning | 79.7 % | independent | Artificial Analysis |
| SciCode | thinking | 50.6 % | independent | Artificial Analysis |
| SciCode | no reasoning | 49.9 % | independent | Artificial Analysis |
external indices
| Intelligence Index v4.1 | thinking | 37.8 points | independent | Artificial Analysis |
| Intelligence Index v4.1 | no reasoning | 27.4 points | independent | Artificial Analysis |
| AA Math Index | thinking | 97.0 points | independent | Artificial Analysis |
| AA Math Index | no reasoning | 55.7 points | independent | Artificial Analysis |
| ECI | 150.6 points | independent | Epoch AI Benchmarking Hub | |
| Vals Index | high effort | 49.5 points | independent | Vals AI |
factuality
| SimpleQA Verified | 67.4 % | independent | Epoch AI Benchmarking Hub 2025-12-17 |
knowledge science
| GPQA Diamond implementation=artificial-analysis | thinking | 89.8 % | independent | Artificial Analysis |
| GPQA Diamond implementation=vals-ai | high effort | 87.9 % | independent | Vals AI |
| GPQA Diamond | 83.2 % | independent | Epoch AI Benchmarking Hub 2025-12-17 | |
| GPQA Diamond implementation=artificial-analysis | no reasoning | 81.2 % | independent | Artificial Analysis |
| Humanity's Last Exam implementation=artificial-analysis | thinking | 34.7 % | independent | Artificial Analysis |
| Humanity's Last Exam implementation=artificial-analysis | no reasoning | 14.1 % | independent | Artificial Analysis |
| MMLU-Pro implementation=artificial-analysis | thinking | 89.0 % | independent | Artificial Analysis |
| MMLU-Pro implementation=vals-ai | high effort | 88.6 % | independent | Vals AI |
| MMLU-Pro implementation=artificial-analysis | no reasoning | 88.2 % | independent | Artificial Analysis |
long context instruction
| AA-LCR | thinking | 66.3 % | independent | Artificial Analysis |
| AA-LCR | no reasoning | 48.0 % | independent | Artificial Analysis |
| IFBench | thinking | 78.0 % | independent | Artificial Analysis |
| IFBench | no reasoning | 55.1 % | independent | Artificial Analysis |
multimodal
| MMMU implementation=vals-ai | high effort | 87.6 % | independent | Vals AI |
professional
| CaseLaw | high effort | 55.8 % | independent | Vals AI |
| CorpFin | high effort | 66.4 % | independent | Vals AI |
| LegalBench | high effort | 86.9 % | independent | Vals AI |
| MedQA | high effort | 95.8 % | independent | Vals AI |
| TaxEval | high effort | 73.9 % | independent | Vals AI |
reasoning math
| AIME year=2025 · implementation=artificial-analysis | thinking | 97.0 % | independent | Artificial Analysis |
| AIME implementation=vals-ai | high effort | 95.6 % | independent | Vals AI |
| AIME year=2025 · implementation=artificial-analysis | no reasoning | 55.7 % | independent | Artificial Analysis |
| ARC-AGI-1older version split=public_eval · model_type=CoT | high effort | 88.3 % | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=semi_private · model_type=CoT | high effort | 84.7 % | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=public_eval · model_type=CoT | medium effort | 67.9 % | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=semi_private · model_type=CoT | medium effort | 57.7 % | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=public_eval · model_type=CoT | low effort | 38.2 % | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=public_eval · model_type=CoT | high effort | 34.0 % | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=semi_private · model_type=CoT | high effort | 33.6 % | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=public_eval · model_type=CoT | minimal effort | 31.9 % | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=semi_private · model_type=CoT | low effort | 29.0 % | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=semi_private · model_type=CoT | minimal effort | 21.5 % | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=public_eval · model_type=CoT | medium effort | 15.3 % | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=semi_private · model_type=CoT | medium effort | 12.8 % | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=semi_private · model_type=CoT | minimal effort | 3.3 % | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=public_eval · model_type=CoT | minimal effort | 2.1 % | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=semi_private · model_type=CoT | low effort | 1.3 % | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=public_eval · model_type=CoT | low effort | 1.3 % | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=public_eval · model_type=CoT | high effort | 0.2 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=semi_private · model_type=CoT | high effort | 0.2 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=semi_private · model_type=CoT | high effort | 0.2 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=public_eval · model_type=CoT | high effort | 0.1 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=public_eval · model_type=CoT | medium effort | 0.1 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=semi_private · model_type=CoT | medium effort | 0.1 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=semi_private · model_type=CoT | medium effort | 0.1 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=public_eval · model_type=CoT | medium effort | 0.1 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=public_eval · model_type=CoT | low effort | 0.0 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=semi_private · model_type=CoT | low effort | 0.0 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=public_eval · model_type=CoT | minimal effort | 0.0 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-2 split=semi_private · model_type=CoT | minimal effort | 0.0 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=semi_private · model_type=CoT | low effort | 0.0 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=public_eval · model_type=CoT | low effort | 0.0 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=semi_private · model_type=CoT | minimal effort | 0.0 usd_per_task | independent | ARC Prize Leaderboard |
| ARC-AGI-1older version split=public_eval · model_type=CoT | minimal effort | 0.0 usd_per_task | independent | ARC Prize Leaderboard |
| FrontierMath | 35.6 % | independent | Epoch AI Benchmarking Hub 2025-12-17 | |
| FrontierMath Tier 4older version | 4.2 % | independent | Epoch AI Benchmarking Hub 2025-12-17 | |
| OTIS Mock AIME 2024–2025 | 92.8 % | independent | Epoch AI Benchmarking Hub 2025-12-17 |
Agent + model results
systems, not bare-model scores
| agent + model mini-SWE-agent + Gemini 3 Flash Preview | SWE-bench bash-only | 75.8 % | unverified | SWE-bench Leaderboard |
| agent + model mini-SWE-agent + Gemini 3 Flash Preview | SWE-bench Verified | 75.8 % | unverified | SWE-bench Leaderboard |
| agent + model Epoch Inspect harness + Gemini 3 Flash Preview | SWE-bench Verified | 75.4 % | independent | Epoch AI Benchmarking Hub |
| agent + model mini-SWE-agent + Gemini 3 Flash Preview | SWE-bench bash-only | 0.4 usd_per_task | unverified | SWE-bench Leaderboard |
| agent + model mini-SWE-agent + Gemini 3 Flash Preview | SWE-bench Verified | 0.4 usd_per_task | unverified | SWE-bench Leaderboard |
| agent + model Junie + Gemini 3 Flash Preview | Terminal-Bench 2.0 | 64.3 % | unverified | Terminal-Bench Leaderboard |
| agent + model terminus-2 + Gemini 3 Flash Preview | Terminal-Bench 2.0 | 51.7 % | community | Terminal-Bench Leaderboard |
| agent + model gemini-cli + Gemini 3 Flash Preview | Terminal-Bench 2.0 | 47.4 % | unverified | Terminal-Bench Leaderboard |
| agent + model Artificial Analysis harness + Gemini 3 Flash Preview | Terminal-Bench Hard | 38.6 % | independent | Artificial Analysis |
| agent + model Artificial Analysis harness + Gemini 3 Flash Preview | Terminal-Bench Hard | 31.8 % | independent | Artificial Analysis |
These scores measure the whole agent system (scaffold, tools, budgets) — they are never merged into the bare model’s numbers.