Head to head
Drex v2.0 scores 78.6 and d3 lite 74.9 on the System One Index. Drex v2.0 is more accurate on 11 of 16 text datasets, d3 lite on 5. Same cases, same request, same scorer.
| Drex v2.0 | d3 lite | |
|---|---|---|
| System One Index | 78.6 | 74.9 |
| English | 77.8 | 74.3 |
| Arabic | 79.6 | 74.7 |
| Vision Index | – | – |
| Calibration error | 1.5% | 7.1% |
| Median latency | 294 ms | – |
| Cost per 1,000 calls | $0.015 | – |
| Drex v2.0 | d3 lite | |
|---|---|---|
| AG News English | 88.0 | 85.5 |
| SST-2 English | 94.0 | 86.0 |
| SST-5 English | 56.0 | 50.5 |
| DAIR Emotion English | 55.5 | 57.0 |
| Banking77 English | 80.5 | 91.0 |
| BoolQ English | 92.5 | 81.0 |
| XNLI English English | 87.3 | 90.7 |
| Typed decisions English | 68.7 | 52.6 |
| XNLI French French | 77.3 | 81.3 |
| XNLI Arabic Arabic | 71.3 | 72.7 |
| MASSIVE scenario Arabic | 63.5 | 58.0 |
| MASSIVE intent Arabic | 82.5 | 77.5 |
| OSACT4 offensive Arabic | 94.5 | 80.5 |
| Tweet sentiment Arabic | 69.5 | 67.0 |
| MIRACL rerank Arabic | 87.0 | 82.5 |
| MIRACL pairwise Arabic | 89.0 | 84.5 |
Model pages: Drex v2.0 · d3 lite. Latency for API models is the median per call through OpenRouter, timed from a cloud server with 8 calls in flight.