Head to head
Drex Max scores 84.1 and Drex v1.5 77.6 on the System One Index. Drex Max is more accurate on 16 of 16 text datasets, Drex v1.5 on 0. Same cases, same request, same scorer.
| Drex Max | Drex v1.5 | |
|---|---|---|
| System One Index | 84.1 | 77.6 |
| English | 82.8 | 77.6 |
| Arabic | 84.5 | 77.0 |
| Vision Index | – | – |
| Calibration error | 6.2% | 10.8% |
| Median latency | 277 ms | 198 ms |
| Cost per 1,000 calls | $0.021 | $0.010 |
| Drex Max | Drex v1.5 | |
|---|---|---|
| AG News English | 90.5 | 90.0 |
| SST-2 English | 96.0 | 94.5 |
| SST-5 English | 59.0 | 53.5 |
| DAIR Emotion English | 65.0 | 53.5 |
| Banking77 English | 90.5 | 87.0 |
| BoolQ English | 93.0 | 89.0 |
| XNLI English English | 95.3 | 88.7 |
| Typed decisions English | 73.0 | 64.8 |
| XNLI French French | 91.3 | 82.0 |
| XNLI Arabic Arabic | 89.3 | 74.0 |
| MASSIVE scenario Arabic | 73.0 | 58.5 |
| MASSIVE intent Arabic | 88.0 | 77.5 |
| OSACT4 offensive Arabic | 93.5 | 90.5 |
| Tweet sentiment Arabic | 69.0 | 63.5 |
| MIRACL rerank Arabic | 89.5 | 89.0 |
| MIRACL pairwise Arabic | 89.5 | 86.0 |
Model pages: Drex Max · Drex v1.5. Latency for API models is the median per call through OpenRouter, timed from a cloud server with 8 calls in flight.