Head to head
Clef Flash scores 80.3 and Drex v2.0 78.6 on the System One Index. Clef Flash is more accurate on 11 of 16 text datasets, Drex v2.0 on 5. Same cases, same request, same scorer.
| Clef Flash | Drex v2.0 | |
|---|---|---|
| System One Index | 80.3 | 78.6 |
| English | 80.2 | 77.8 |
| Arabic | 80.5 | 79.6 |
| Vision Index | 78.0 | – |
| Calibration error | 3.5% | 1.5% |
| Median latency | 267 ms | 294 ms |
| Cost per 1,000 calls | $0.044 | $0.015 |
| Clef Flash | Drex v2.0 | |
|---|---|---|
| AG News English | 90.0 | 88.0 |
| SST-2 English | 95.0 | 94.0 |
| SST-5 English | 57.0 | 56.0 |
| DAIR Emotion English | 57.0 | 55.5 |
| Banking77 English | 98.5 | 80.5 |
| BoolQ English | 88.5 | 92.5 |
| XNLI English English | 84.7 | 87.3 |
| Typed decisions English | 70.5 | 68.7 |
| XNLI French French | 80.0 | 77.3 |
| XNLI Arabic Arabic | 68.7 | 71.3 |
| MASSIVE scenario Arabic | 89.5 | 63.5 |
| MASSIVE intent Arabic | 90.5 | 82.5 |
| OSACT4 offensive Arabic | 89.5 | 94.5 |
| Tweet sentiment Arabic | 47.0 | 69.5 |
| MIRACL rerank Arabic | 88.0 | 87.0 |
| MIRACL pairwise Arabic | 90.0 | 89.0 |
| Clef Flash | Drex v2.0 | |
|---|---|---|
| SST-2, as an image | 95.5 | – |
| Arabic sentiment, as an image | 33.5 | – |
| Oxford-IIIT Pet | 85.0 | – |
| VQAv2 yes/no | 88.0 | – |
| CountBench | 88.0 | – |
Model pages: Clef Flash · Drex v2.0. Latency for API models is the median per call through OpenRouter, timed from a cloud server with 8 calls in flight.