~/system-one-bench / compare/google-gemini-decision-1-vs-vllm-semantic-router-d3-nano
Head to head
Gemini Decision 1 scores 81.6 and d3 nano 78.0 on the System One Index. Gemini Decision 1 is more accurate on 11 of 16 text datasets, d3 nano on 3, and 2 are tied. Same cases, same request, same scorer.
| Gemini Decision 1 | d3 nano | |
|---|---|---|
| System One Index | 81.6 | 78.0 |
| English | 79.2 | 77.9 |
| Arabic | 84.5 | 76.9 |
| Vision Index | 89.2 | – |
| Calibration error | 3.6% | 7.0% |
| Median latency | – | – |
| Cost per 1,000 calls | $0.036 | – |
| Gemini Decision 1 | d3 nano | |
|---|---|---|
| AG News English | 88.5 | 88.0 |
| SST-2 English | 96.5 | 89.5 |
| SST-5 English | 59.0 | 55.5 |
| DAIR Emotion English | 59.5 | 59.5 |
| Banking77 English | 79.5 | 93.0 |
| BoolQ English | 89.0 | 88.0 |
| XNLI English English | 87.3 | 87.3 |
| Typed decisions English | 74.0 | 62.6 |
| XNLI French French | 81.3 | 86.0 |
| XNLI Arabic Arabic | 78.0 | 82.0 |
| MASSIVE scenario Arabic | 72.5 | 61.5 |
| MASSIVE intent Arabic | 87.0 | 81.0 |
| OSACT4 offensive Arabic | 95.5 | 79.0 |
| Tweet sentiment Arabic | 78.0 | 71.0 |
| MIRACL rerank Arabic | 89.5 | 77.5 |
| MIRACL pairwise Arabic | 91.0 | 86.5 |
| Gemini Decision 1 | d3 nano | |
|---|---|---|
| SST-2, as an image | 94.5 | – |
| Arabic sentiment, as an image | 75.0 | – |
| Oxford-IIIT Pet | 98.0 | – |
| VQAv2 yes/no | 86.0 | – |
| CountBench | 92.5 | – |
Model pages: Gemini Decision 1 · d3 nano. Latency for API models is the median per call through OpenRouter, timed from a cloud server with 8 calls in flight.