Head to head
Drex v1.5 scores 77.6 and Drex v2.0 78.6 on the System One Index. Drex v1.5 is more accurate on 7 of 16 text datasets, Drex v2.0 on 9. Same cases, same request, same scorer.
| Drex v1.5 | Drex v2.0 | |
|---|---|---|
| System One Index | 77.6 | 78.6 |
| English | 77.6 | 77.8 |
| Arabic | 77.0 | 79.6 |
| Vision Index | – | – |
| Calibration error | 10.8% | 1.5% |
| Median latency | 198 ms | 294 ms |
| Cost per 1,000 calls | $0.010 | $0.015 |
| Drex v1.5 | Drex v2.0 | |
|---|---|---|
| AG News English | 90.0 | 88.0 |
| SST-2 English | 94.5 | 94.0 |
| SST-5 English | 53.5 | 56.0 |
| DAIR Emotion English | 53.5 | 55.5 |
| Banking77 English | 87.0 | 80.5 |
| BoolQ English | 89.0 | 92.5 |
| XNLI English English | 88.7 | 87.3 |
| Typed decisions English | 64.8 | 68.7 |
| XNLI French French | 82.0 | 77.3 |
| XNLI Arabic Arabic | 74.0 | 71.3 |
| MASSIVE scenario Arabic | 58.5 | 63.5 |
| MASSIVE intent Arabic | 77.5 | 82.5 |
| OSACT4 offensive Arabic | 90.5 | 94.5 |
| Tweet sentiment Arabic | 63.5 | 69.5 |
| MIRACL rerank Arabic | 89.0 | 87.0 |
| MIRACL pairwise Arabic | 86.0 | 89.0 |
Model pages: Drex v1.5 · Drex v2.0. Latency for API models is the median per call through OpenRouter, timed from a cloud server with 8 calls in flight.