| Benchmark | Gemma 4 E2B | GPT-5 |
|---|---|---|
| AA Agentic Index | 7.4 | 49.7 |
| AA Intelligence | 10 | 35.3 |
| AA-LCR | 17 | 76.3 |
| AA-Omniscience | -23.6 | -8.7 |
| AIME25 no tools | 26.3 | 94.6 |
| Artificial Analysis Coding Index | 8.3 | 38.9 |
| critpt | 0 | 5.7 |
| gdpval | 0 | 32.5 |
| GPQA Diamond | 43.4 | 87.3 |
| HLE | 4.8 | 28.5 |
| IFBench | 38 | 73.1 |
| LiveCodeBench v6 | 44 | 87 |
| LiveCodeBenchV6 no tools | 54.9 | 87 |
| MMLU-Pro | 60 | 87.5 |
| MMMU-Pro | 44.6 | 78.4 |
| ocrbench_v2 | 44.7 | 55.5 |
| OmniScience Accuracy | 8.2 | 40.3 |
| OmniScience Non-Hallucination | 67.6 | 21 |
| scicode | 20.9 | 43 |
| Tau2 retail | 18.9 | 81.1 |
| TauBench V3 - Banking | 4.5 | 22.1 |
| Terminal-Bench 2.1 | 0.4 | 35.2 |
| Terminal-Bench Hard | 3 | 37.9 |
| τ²-Bench | 24.5 | 85 |
| τ²-Bench Telecom (AA run) | 22.2 | 86.5 |
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.