Google · model
Gemini 3.8 Flash benchmark scores
As of 2026-11-22, Gemini 3.8 Flash (Google) has measured scores on 65 benchmarks and estimated scores on 121 more.
| Benchmark | Score | Source |
|---|---|---|
| AA-GPQA Diamond | 95.3% | measured |
| AA-HLE | 47.8% | measured |
| AA-SciCode | 56.6% | measured |
| AIME26 | 95.9% | estimated ± 1.2 pp, medium confidence |
| Artificial Analysis Intelligence Index | 30.9% | measured |
| CritPt | 18.3% | measured |
| GDPval-AA | 46.8% | measured |
| GPQA Diamond | 96.4% | estimated ± 1.4 pp, high confidence |
| HLE | 57.2% | estimated ± 2.2 pp, high confidence |
| AA-LCR | 81.3% | measured |
| LiveCodeBench v6 | 90.7% | estimated ± 2.2 pp, low confidence |
| AA-Omniscience Accuracy | 50.5% | measured |
| SciCode | 55.8% | estimated ± 2.9 pp, low confidence |
| Terminal-Bench 2.1 | 89.4% | measured |
| BrowseComp | 91.6% | estimated ± 1.9 pp, high confidence |
| HLE w/ tools | 45.5% | estimated ± 3.1 pp, medium confidence |
| LongBench v2 | 66.8% | estimated ± 5.5 pp, medium confidence |
| τ²-bench results | 93.2% | estimated ± 6.5 pp, low confidence |
| AA Coding Index | 76.3% | measured |
| AA-MMMU-Pro | 85.6% | measured |
| APEX-Agents | 39.2% | estimated ± 9.4 pp, low confidence |
| APEX-Agents-AA | 38.7% | estimated ± 2.9 pp, low confidence |
| BioMysteryBench (human-difficult) | 56.5% | measured |
| DeepSearchQA | 91.9% | estimated ± 4.5 pp, high confidence |
| SWE-bench Verified | 81.9% | estimated ± 4.4 pp, high confidence |
| AutomationBench | 39.2% | estimated ± 9.5 pp, low confidence |
| BFCL v4 | 76.4% | estimated ± 2.7 pp, medium confidence |
| CyberGym | 85.6% | estimated ± 4.5 pp, high confidence |
| JobBench | 64.6% | estimated ± 4.9 pp, medium confidence |
| SWE-bench Pro | 73.4% | estimated ± 4.5 pp, high confidence |
| WideResearch | 84.9% | estimated ± 3.3 pp, medium confidence |
| AA Agentic Index | 41.1% | measured |
| MMLU-Pro | 87.8% | estimated ± 4.8 pp, medium confidence |
| FailureSensorIQ | 69.0% | measured |
| GPQA | 93.6% | estimated ± 0.9 pp, high confidence |
| MMLU | 90.5% | estimated ± 0.9 pp, medium confidence |
| Vals GPQA Diamond | 94.4% | measured |
| Vals LiveCodeBench | 89.5% | measured |
| Vals MMLU-Pro | 90.2% | measured |
| MMMLU | 90.6% | estimated ± 0.7 pp, low confidence |
| MMMU | 89.7% | estimated ± 2.7 pp, medium confidence |
| Chess Puzzles | 58.0% | measured |
| AA MMLU-Pro | 88.9% | estimated ± 0.7 pp, medium confidence |
| Gert Labs | 67.0% | measured |
| AA-AnalystAgent | 46.3% | measured |
| AA EnterpriseOps-Gym | 35.4% | estimated ± 7.2 pp, low confidence |
| AA Harvey LAB v1.0 | 92.4% | estimated ± 1.2 pp, medium confidence |
| AA Tau3 Banking | 47.1% | estimated ± 2.4 pp, low confidence |
| AA Terminal-Bench 2.1 | 87.6% | measured |
| ApprenticeBench | 24.0% | measured |
| ARC-AGI-1 | 98.5% | measured |
| ARC-AGI-2 | 89.2% | measured |
| BabaIsBench | 100.0% | measured |
| CursorBench 3.1 | 60.3% | estimated ± 3.4 pp, medium confidence |
| CursorBench 3.2 | 69.2% | measured |
| EBR-bench | 18.9% | estimated ± 9.3 pp, medium confidence |
| Terminal-Bench 3.0 | 21.4% | estimated ± 3.0 pp, high confidence |
| FrontierCode 1.1 Main | 52.8% | estimated ± 1.6 pp, medium confidence |
| FrontierMath v2 (Tier 4) | 22.0% | measured |
| FrontierMath v2 (Tiers 1-3) | 68.4% | measured |
| FrontierSWE v2 | 19.6% | measured |
| Furniture Assembly | 31.7% | measured |
| MaxxBench | 53.4% | measured |
| MazeBench | 4.0% | measured |
| MirrorCode | 38.9% | measured |
| Mystery Game Puzzles | 47.0% | measured |
| NYT Connections | 97.4% | measured |
| OfficeQA Pro | 67.6% | estimated ± 6.8 pp, medium confidence |
| OSWorld-Verified | 81.3% | estimated ± 2.3 pp, medium confidence |
| ParseBench-1 | 71.2% | measured |
| Terminal-Bench 4.1 | 18.2% | measured |
| Terminal-Bench Hard | 57.5% | estimated ± 2.8 pp, medium confidence |
| Vals SWE-bench | 80.0% | measured |
| Terminal-Bench 2.1 (Vals AI) | 81.3% | measured |
| VulcanBench v3 | 78.2% | estimated ± 3.9 pp, medium confidence |
| CWE-Bench-AA | 52.5% | measured |
| AA Cyber Index | 25.1% | measured |
| DeepsecBench-AA | 14.5% | measured |
| Harvey LAB-AA v1.1 | 7.4% | estimated ± 4.0 pp, low confidence |
| HealthBench Professional (length-adjusted, Artificial Analysis) | 55.2% | estimated ± 4.3 pp, low confidence |
| HealthBench Professional (raw, Artificial Analysis) | 64.4% | estimated ± 0.8 pp, low confidence |
| AA ITBench | 52.5% | measured |
| Bug Hunt Bench | 18.0% | measured |
| Chartography (tools) | 86.7% | estimated ± 4.1 pp, low confidence |
| CursorBench 4.0 | 42.0% | estimated ± 2.0 pp, high confidence |
| CWE-bench v1 | 57.4% | estimated ± 5.1 pp, medium confidence |
| DeepSWE | 73.8% | measured |
| DRACO | 86.9% | estimated ± 0.7 pp, low confidence |
| FrontierCode 2.0 | 24.3% | estimated ± 4.8 pp, medium confidence |
| FrontierPhysics | 0.0% | measured |
| HLE w/o tools | 51.4% | estimated ± 4.9 pp, high confidence |
| OSWorld 2.0 | 59.0% | measured |
| OSWorld 2.1 | 76.3% | estimated ± 3.6 pp, high confidence |
| OSWorld 2.1 Offline | 50.3% | estimated ± 1.9 pp, medium confidence |
| PostTrainBench v1.1 | 24.1% | estimated ± 3.7 pp, high confidence |
| ProgramBench | 88.0% | estimated ± 6.0 pp, low confidence |
| SpireClimb | 16.8% | estimated ± 7.2 pp, medium confidence |
| StealthMark | 47.9% | measured |
| SWE-bench Multimodal | 56.9% | estimated ± 5.6 pp, low confidence |
| Terminal-Bench 4.0 | 19.1% | measured |
| Terminal-Bench-Science 0.1 | 0.0% | estimated ± 10.7 pp, low confidence |
| Toolathlon-Verified | 79.3% | estimated ± 1.0 pp, medium confidence |
| VulcanBench Frontier v4 | 90.1% | estimated ± 2.3 pp, medium confidence |
| AA AutomationBench | 51.2% | estimated ± 5.5 pp, medium confidence |
| GDP.pdf | 18.9% | estimated ± 3.1 pp, high confidence |
| AA Terminal-Bench 4.0 | 17.8% | estimated ± 6.2 pp, medium confidence |
| ARC-AGI-3 | 10.4% | measured |
| ARC-AGI-3 Provider Adapter | 35.0% | measured |
| cursorBench41 | 44.8% | estimated ± 3.7 pp, high confidence |
| CWE-bench | 47.2% | estimated ± 4.3 pp, medium confidence |
| MKA | 35.3% | measured |
| SkillsBench official leaderboard (BenchFlow) | 58.6% | estimated ± 6.1 pp, medium confidence |
| BinaryAudit | 40.5% | estimated ± 8.9 pp, low confidence |
| OTelBench | 22.5% | estimated ± 6.6 pp, low confidence |
| Vibe Code Bench | 76.2% | estimated ± 10.0 pp, low confidence |
| FACTS Multimodal | 57.6% | measured |
| Hermes Bench | 62.2% | measured |
| Hermes Index | 38.1% | measured |
| ParseBench-2 | 76.7% | measured |
| CharXiv | 91.2% | estimated ± 2.4 pp, high confidence |
| CharXiv w/o tools | 86.2% | measured |
| ExploitBench | 49.7% | estimated ± 1.9 pp, medium confidence |
| ExploitGym | 19.5% | estimated ± 2.7 pp, high confidence |
| AI2D_TEST | 94.2% | estimated ± 2.0 pp, low confidence |
| C-Eval | 94.8% | estimated ± 1.0 pp, low confidence |
| CountBench | 100.0% | estimated ± 2.9 pp, low confidence |
| DynaMath | 85.5% | estimated ± 3.0 pp, low confidence |
| ERQA | 73.7% | estimated ± 2.9 pp, medium confidence |
| HMMT Feb 2026 | 91.8% | estimated ± 1.5 pp, medium confidence |
| HMMT Nov 2025 | 95.4% | estimated ± 1.5 pp, low confidence |
| MathVision | 94.8% | estimated ± 3.1 pp, medium confidence |
| MCP Atlas | 80.8% | estimated ± 5.8 pp, low confidence |
| MMAnswerBench | 91.3% | estimated ± 2.4 pp, low confidence |
| MMLU-Redux | 97.3% | estimated ± 0.8 pp, low confidence |
| MMMU-Pro | 85.0% | estimated ± 1.5 pp, medium confidence |
| NL2Repo | 45.5% | estimated ± 4.7 pp, high confidence |
| PostTrainBench v1 | 24.4% | estimated ± 3.6 pp, high confidence |
| QwenClawBench | 93.2% | estimated ± 6.0 pp, low confidence |
| RealWorldQA | 87.9% | estimated ± 1.8 pp, medium confidence |
| ScreenSpot Pro | 90.2% | estimated ± 10.1 pp, low confidence |
| SuperGPQA | 99.9% | estimated ± 9.3 pp, low confidence |
| Terminal-Bench 2.0 | 79.8% | estimated ± 5.8 pp, low confidence |
| Toolathlon | 56.7% | estimated ± 1.7 pp, medium confidence |
| V* | 100.0% | estimated ± 7.5 pp, low confidence |
| Video-MME (w/o subtitle) | 90.9% | estimated ± 1.7 pp, low confidence |
| Video-MME (with subtitle) | 91.6% | estimated ± 3.7 pp, low confidence |
| AA LiveCodeBench | 91.6% | estimated ± 2.6 pp, medium confidence |
| HealthBench Hard | 32.2% | estimated ± 8.6 pp, low confidence |
| MedXpertQA (MM) | 77.8% | estimated ± 6.5 pp, low confidence |
| React Native Evals | 83.3% | estimated ± 3.2 pp, medium confidence |
| ResearchClawBench | 20.3% | estimated ± 1.9 pp, high confidence |
| PostTrainBench | 33.1% | estimated ± 2.4 pp, high confidence |
| DAYJOB | 2.6% | measured |
| DAYJOB Healthcare | 1.2% | measured |
| Finance Agent v2 | 61.4% | measured |
| BioMysteryBench (human-solvable) | 88.8% | measured |
| FrontierCode 1.1 Extended | 62.3% | estimated ± 1.6 pp, medium confidence |
| HealthBench (raw) | 55.6% | estimated ± 6.0 pp, medium confidence |
| HealthBench (length-adjusted) | 53.9% | estimated ± 0.9 pp, medium confidence |
| HealthBench Professional | 59.4% | estimated ± 2.1 pp, medium confidence |
| HealthBench Professional (raw) | 62.2% | estimated ± 2.3 pp, medium confidence |
| HLE-Verified | 54.9% | measured |
| LABBench2 | 86.2% | measured |
| ScarfBench | 66.0% | measured |
| BlenderWorld (text-only) | 77.7% | estimated ± 2.8 pp, medium confidence |
| Cookie-Bench | 87.3% | estimated ± 1.3 pp, medium confidence |
| STEMBench | 2.6% | estimated ± 0.2 pp, medium confidence |
| SimpleQA | 36.3% | estimated ± 7.8 pp, low confidence |
| Agents' Last Exam | 39.0% | estimated ± 9.7 pp, medium confidence |
| BabyVision w/ Python | 90.6% | estimated ± 3.7 pp, low confidence |
| OpenHarmony Bench v1.0 | 58.9% | estimated ± 2.4 pp, high confidence |
| OpenHarmony Bench v1.2 | 52.8% | estimated ± 11.0 pp, low confidence |
| SEC-Bench Pro | 85.2% | estimated ± 5.6 pp, medium confidence |
| BabyVision | 70.6% | estimated ± 7.4 pp, low confidence |
| PerceptionBench | 65.2% | estimated ± 3.8 pp, medium confidence |
| ZeroBench | 25.9% | estimated ± 7.1 pp, low confidence |
| MRCRv2 | 85.9% | estimated ± 11.3 pp, low confidence |
| LVBench | 87.1% | measured |
| MMLongBench-Doc | 61.5% | estimated ± 2.6 pp, low confidence |
| SWE-Marathon | 41.3% | estimated ± 2.1 pp, low confidence |
| BioBasicsBench | 38.7% | estimated ± 2.8 pp, medium confidence |
| LifeVal | 77.8% | estimated ± 2.3 pp, low confidence |
| MMMU-Pro w/ Python | 87.8% | estimated ± 2.0 pp, medium confidence |
| ExploitBench Internal Port | 5.8% | estimated ± 1.4 pp, medium confidence |
| MathVision w/ Python | 97.0% | estimated ± 1.3 pp, low confidence |
| SWE-Atlas Codebase QnA | 61.4% | estimated ± 4.8 pp, high confidence |