Head to head
Mercury Decide scores 82.4 and Kev 4B 76.8 on the System One Index. Mercury Decide is more accurate on 12 of 16 text datasets, Kev 4B on 3, and 1 are tied. Same cases, same request, same scorer.
| Mercury Decide | Kev 4B | |
|---|---|---|
| System One Index | 82.4 | 76.8 |
| English | 83.7 | 78.3 |
| Arabic | 80.8 | 74.8 |
| Vision Index | 79.4 | – |
| Calibration error | 13.1% | 5.1% |
| Median latency | 276 ms | 552 ms |
| Cost per 1,000 calls | Free | $0.010 |
| Mercury Decide | Kev 4B | |
|---|---|---|
| AG News English | 92.0 | 90.5 |
| SST-2 English | 95.5 | 92.5 |
| SST-5 English | 52.0 | 52.0 |
| DAIR Emotion English | 91.5 | 54.5 |
| Banking77 English | 86.0 | 88.0 |
| BoolQ English | 90.5 | 91.5 |
| XNLI English English | 91.3 | 90.7 |
| Typed decisions English | 71.1 | 66.3 |
| XNLI French French | 82.7 | 79.3 |
| XNLI Arabic Arabic | 76.0 | 72.7 |
| MASSIVE scenario Arabic | 72.0 | 57.0 |
| MASSIVE intent Arabic | 85.0 | 76.5 |
| OSACT4 offensive Arabic | 92.0 | 88.5 |
| Tweet sentiment Arabic | 62.5 | 63.0 |
| MIRACL rerank Arabic | 88.0 | 79.5 |
| MIRACL pairwise Arabic | 90.0 | 86.5 |
| Mercury Decide | Kev 4B | |
|---|---|---|
| SST-2, as an image | 93.0 | – |
| Arabic sentiment, as an image | 39.0 | – |
| Oxford-IIIT Pet | 84.5 | – |
| VQAv2 yes/no | 87.0 | – |
| CountBench | 93.5 | – |
Model pages: Mercury Decide · Kev 4B. Latency for API models is the median per call through OpenRouter, timed from one machine.