| Benchmark | Gemma 4 26B A4B | GPT-4o |
|---|---|---|
| AA Agentic Index | 32.1 | 9.7 |
| AA Intelligence | 26.1 | 11.2 |
| AA-LCR | 61.7 | 55 |
| AA-Omniscience | -50.8 | -10.5 |
| AI2D | 88.3 | 94.2 |
| AIME 2025 | 90 | 11.7 |
| arena_vision | 1238 | 1162 |
| Artificial Analysis Coding Index | 39.3 | 24.2 |
| C-Eval | 82.5 | 76 |
| critpt | 0 | 0 |
| gdpval | 25.7 | 0 |
| GPQA Diamond | 82.3 | 70.1 |
| GSM8K | 77 | 95.6 |
| HLE | 19.3 | 5.3 |
| humaneval | 50 | 90.6 |
| IFBench | 72.4 | 36 |
| ifeval | 96.4 | 84.3 |
| LiveCodeBench | 79.8 | 38.3 |
| LiveCodeBench v6 | 77.1 | 30.9 |
| mathvision | 82.4 | 30.4 |
| MathVista | 79.4 | 63.8 |
| mmlu | 77.8 | 88.1 |
| mmlu_redux | 92.7 | 88 |
| MMLU-Pro | 85.2 | 74.7 |
| mmmlu | 86.3 | 81.4 |
| MMMU | 78.4 | 72.2 |
| MMMU-Pro | 73.8 | 59.9 |
| OmniScience Accuracy | 19.1 | 23.7 |
| OmniScience Non-Hallucination | 13.6 | 62.1 |
| OpenBookQA | 48.8 | 96.8 |
| RealWorldQA | 72.2 | 75.4 |
| scicode | 40.3 | 33.4 |
| supergpqa | 61.4 | 42.4 |
| SWE-bench Verified | 57.4 | 38.8 |
| Terminal-Bench Hard | 25 | 8.3 |
| τ²-Bench Telecom (AA run) | 43.6 | 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.