| Benchmark | Gemma 4 26B A4B | Opus 5 |
|---|---|---|
| AA Agentic Index | 32.1 | 59.2 |
| AA Intelligence | 26.1 | 63.1 |
| AA-LCR | 61.7 | 78.7 |
| AA-Omniscience | -50.8 | 37.1 |
| arena_elo | 1438 | 1487 |
| arena_text_factuality | 1445 | 1481 |
| Artificial Analysis Coding Index | 39.3 | 78 |
| browsecomp | 26.3 | 90.8 |
| coding_arena_elo | 1361 | 1662 |
| critpt | 0 | 29.1 |
| gdpval | 25.7 | 67.2 |
| GDPval-AA v2 | 807 | 1861 |
| GPQA Diamond | 82.3 | 93.7 |
| HLE | 19.3 | 64.7 |
| HLE (with tools) | 17.2 | 64.7 |
| MCP Atlas | 50 | 85.8 |
| MMMU-Pro | 73.8 | 84.7 |
| OmniScience Accuracy | 19.1 | 60.9 |
| OmniScience Non-Hallucination | 13.6 | 40.5 |
| scicode | 40.3 | 55.7 |
| SWE-bench Multilingual | 43.4 | 89.5 |
| SWE-bench Pro | 13.8 | 79.2 |
| SWE-bench Verified | 57.4 | 96 |
| TauBench V3 - Banking | 12 | 44.7 |
| Terminal-Bench 2.1 | 39 | 89.1 |
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.