| Benchmark | Gemma 4 E2B | GPT-4o |
|---|---|---|
| AA Agentic Index | 7.4 | 9.7 |
| AA Intelligence | 10 | 11.2 |
| AA-LCR | 17 | 55 |
| AA-Omniscience | -23.6 | -10.5 |
| Artificial Analysis Coding Index | 8.3 | 24.2 |
| chartqa | 43.5 | 88.1 |
| critpt | 0 | 0 |
| gdpval | 0 | 0 |
| GPQA Diamond | 43.4 | 70.1 |
| HLE | 4.8 | 5.3 |
| IFBench | 38 | 36 |
| ifeval | 82.9 | 84.3 |
| LiveCodeBench v6 | 44 | 30.9 |
| MATH500 (Pass@1) | 64 | 74.6 |
| mathvision | 52.4 | 30.4 |
| MathVista | 62.1 | 63.8 |
| MMLU-Pro | 60 | 74.7 |
| mmmlu | 67.4 | 81.4 |
| MMMU-Pro | 44.6 | 59.9 |
| MMStar | 57.9 | 63.9 |
| OCRBench | 0.7 | 806 |
| OmniScience Accuracy | 8.2 | 23.7 |
| OmniScience Non-Hallucination | 67.6 | 62.1 |
| RealWorldQA | 56.2 | 75.4 |
| scicode | 20.9 | 33.4 |
| Tau2 retail | 18.9 | 63.4 |
| Terminal-Bench Hard | 3 | 8.3 |
| τ²-Bench Telecom (AA run) | 22.2 | 28.9 |
Best tracked score per model per benchmark (default configuration; source-attributed). ↓ marks lower-is-better metrics. Open either model for its full surface, provenance and pricing.