| Benchmark | Gemma 4 26B A4B | Gemma 4 E2B |
|---|---|---|
| AA Agentic Index | 32.1 | 7.4 |
| AA Intelligence | 26.1 | 10 |
| AA-LCR | 61.7 | 17 |
| AA-Omniscience | -50.8 | -23.6 |
| aime_2026 | 88.3 | 37.5 |
| Artificial Analysis Coding Index | 39.3 | 8.3 |
| BIG-Bench Extra Hard | 64.8 | 21.9 |
| Claw-Eval Avg | 58.8 | 53.1 |
| Codeforces ELO | 1718 | 633 |
| critpt | 0 | 0 |
| gdpval | 25.7 | 0 |
| GPQA Diamond | 82.3 | 43.4 |
| HallusionBench | 66.1 | 0.6 |
| HLE | 19.3 | 4.8 |
| IFBench | 72.4 | 38 |
| ifeval | 96.4 | 82.9 |
| LiveCodeBench v6 | 77.1 | 44 |
| mathvision | 82.4 | 52.4 |
| MathVista | 79.4 | 62.1 |
| MedXPertQA MM | 58.1 | 23.5 |
| MMLU-Pro | 85.2 | 60 |
| mmmlu | 86.3 | 67.4 |
| MMMU-Pro | 73.8 | 44.6 |
| mrcr_v2_8needle_128k_average | 44.1 | 19.1 |
| OmniDocBench 1.5 (average edit distance, lower is better) | 0.1 | 0.3 |
| OmniScience Accuracy | 19.1 | 8.2 |
| OmniScience Non-Hallucination | 13.6 | 67.6 |
| PinchBench | 74.7 | 44.2 |
| RealWorldQA | 72.2 | 56.2 |
| scicode | 40.3 | 20.9 |
| TauBench V3 - Banking | 12 | 4.5 |
| Terminal-Bench 2.1 | 39 | 0.4 |
| Terminal-Bench Hard | 25 | 3 |
| τ²-Bench | 68.2 | 24.5 |
| τ²-Bench Telecom (AA run) | 43.6 | 22.2 |
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.