Google · model
Gemini 4 Carbon benchmark scores
As of 2026-10-29, Gemini 4 Carbon (Google) has measured scores on 16 benchmarks and estimated scores on 71 more.
| Benchmark | Score | Source |
|---|---|---|
| AA-GPQA Diamond | 93.7% | estimated ± 6.6 pp, low confidence |
| AA-HLE | 60.4% | measured |
| AA-SciCode | 64.2% | measured |
| CritPt | 36.0% | measured |
| GDPval-AA | 63.3% | estimated ± 2.9 pp, medium confidence |
| GPQA Diamond | 94.2% | estimated ± 2.3 pp, low confidence |
| HLE | 60.4% | estimated ± 9.7 pp, low confidence |
| AA-LCR | 79.3% | measured |
| LiveCodeBench v6 | 100.0% | estimated ± 4.9 pp, medium confidence |
| SciCode | 63.6% | estimated ± 3.1 pp, medium confidence |
| Terminal-Bench 2.1 | 87.4% | estimated ± 5.2 pp, low confidence |
| BrowseComp | 96.6% | estimated ± 3.0 pp, low confidence |
| HLE w/ tools | 70.9% | estimated ± 5.2 pp, low confidence |
| LongBench v2 | 62.0% | estimated ± 5.8 pp, low confidence |
| AA Coding Index | 78.0% | estimated ± 3.0 pp, high confidence |
| BioMysteryBench (human-difficult) | 50.3% | estimated ± 6.8 pp, low confidence |
| SWE-bench Verified | 100.0% | estimated ± 6.4 pp, low confidence |
| AutomationBench | 53.9% | measured |
| CyberGym | 87.9% | estimated ± 4.6 pp, medium confidence |
| JobBench | 69.1% | estimated ± 12.9 pp, low confidence |
| SWE-bench Pro | 84.6% | estimated ± 5.1 pp, medium confidence |
| AA Agentic Index | 58.2% | estimated ± 8.0 pp, low confidence |
| MMLU-Pro | 86.2% | estimated ± 4.1 pp, medium confidence |
| Artificial Analysis Intelligence Index | 59.4% | measured |
| GPQA | 93.7% | estimated ± 6.3 pp, low confidence |
| AA-Omniscience Accuracy | 58.3% | measured |
| MMMLU | 93.2% | estimated ± 2.6 pp, low confidence |
| AA MMLU-Pro | 90.0% | estimated ± 0.7 pp, medium confidence |
| AA-AnalystAgent | 54.8% | estimated ± 2.8 pp, medium confidence |
| AA Tau3 Banking | 50.2% | estimated ± 3.8 pp, low confidence |
| AA Terminal-Bench 2.1 | 87.5% | estimated ± 4.8 pp, low confidence |
| ARC-AGI-1 | 95.6% | estimated ± 7.6 pp, medium confidence |
| ARC-AGI-2 | 92.1% | estimated ± 14.6 pp, low confidence |
| CursorBench 3.1 | 79.0% | estimated ± 3.4 pp, low confidence |
| CursorBench 3.2 | 76.1% | estimated ± 4.1 pp, low confidence |
| FrontierCode 1.1 Main | 51.8% | estimated ± 3.8 pp, low confidence |
| FrontierSWE v2 | 61.2% | measured |
| MazeBench | 25.5% | estimated ± 3.0 pp, medium confidence |
| Vals GPQA Diamond | 93.3% | estimated ± 5.3 pp, low confidence |
| Vals MMLU-Pro | 88.6% | estimated ± 2.1 pp, medium confidence |
| Vals SWE-bench | 99.5% | estimated ± 4.9 pp, medium confidence |
| Terminal-Bench 2.1 (Vals AI) | 80.9% | estimated ± 9.4 pp, low confidence |
| AA AutomationBench | 77.8% | measured |
| CWE-Bench-AA | 61.9% | estimated ± 6.0 pp, low confidence |
| AA Cyber Index | 46.3% | estimated ± 9.4 pp, medium confidence |
| DeepsecBench-AA | 39.4% | estimated ± 5.3 pp, low confidence |
| GDP.pdf | 28.4% | measured |
| AA ITBench | 54.3% | estimated ± 4.3 pp, medium confidence |
| AA Terminal-Bench 4.0 | 65.2% | measured |
| Bug Hunt Bench | 42.2% | estimated ± 7.6 pp, medium confidence |
| CursorBench 4.0 | 55.7% | estimated ± 2.7 pp, high confidence |
| CWE-bench v1 | 71.0% | measured |
| DeepSWE | 73.3% | estimated ± 4.5 pp, medium confidence |
| FrontierCode 2.0 | 38.6% | estimated ± 4.9 pp, high confidence |
| GraphWalks BFS 256K–1M | 96.2% | measured |
| HLE w/o tools | 63.8% | estimated ± 4.6 pp, medium confidence |
| OSWorld 2.0 | 53.0% | estimated ± 9.2 pp, low confidence |
| OSWorld 2.1 | 79.9% | estimated ± 2.8 pp, low confidence |
| PostTrainBench v1.1 | 44.2% | estimated ± 4.3 pp, medium confidence |
| Terminal-Bench 4.0 | 64.6% | estimated ± 5.7 pp, medium confidence |
| Toolathlon-Verified | 78.5% | estimated ± 1.3 pp, low confidence |
| OSWorld 2.1 Offline | 81.5% | measured |
| ARC-AGI-3 | 48.0% | estimated ± 7.7 pp, low confidence |
| ARC-AGI-3 Provider Adapter | 98.6% | estimated ± 7.5 pp, low confidence |
| Hermes Bench | 73.8% | estimated ± 0.3 pp, low confidence |
| Hermes Index | 61.0% | estimated ± 2.6 pp, low confidence |
| ExploitBench | 100.0% | estimated ± 10.7 pp, low confidence |
| ExploitGym | 53.5% | measured |
| MMLU-Redux | 98.2% | estimated ± 0.8 pp, low confidence |
| NL2Repo | 74.7% | estimated ± 7.4 pp, low confidence |
| SuperGPQA | 100.0% | estimated ± 9.3 pp, low confidence |
| HealthBench Hard | 34.4% | estimated ± 8.6 pp, low confidence |
| React Native Evals | 82.7% | estimated ± 4.7 pp, low confidence |
| Finance Agent v2 | 63.5% | estimated ± 4.5 pp, low confidence |
| BioMysteryBench (human-solvable) | 89.6% | estimated ± 0.6 pp, low confidence |
| HealthBench (raw) | 65.4% | estimated ± 6.0 pp, medium confidence |
| HealthBench (length-adjusted) | 60.7% | estimated ± 3.3 pp, medium confidence |
| HealthBench Professional | 65.4% | estimated ± 4.5 pp, medium confidence |
| HealthBench Professional (raw) | 75.2% | estimated ± 4.5 pp, high confidence |
| HLE-Verified | 54.4% | estimated ± 6.8 pp, low confidence |
| LABBench2 | 88.5% | estimated ± 2.4 pp, low confidence |
| STEMBench | 2.7% | estimated ± 0.2 pp, medium confidence |
| Agents' Last Exam | 51.7% | estimated ± 9.6 pp, low confidence |
| OpenHarmony Bench | 58.9% | estimated ± 2.4 pp, medium confidence |
| SEC-Bench Pro | 88.2% | measured |
| MRCRv2 | 85.7% | estimated ± 11.3 pp, low confidence |
| MMLongBench-Doc | 61.4% | estimated ± 2.6 pp, low confidence |