~/system-one-bench / compare/openai-gpt-6-luna-decisions-vs-patrick-18-dexy-decision-9b
Head to head
GPT-6 Luna Decisions scores 76.0 and Dexy-Decision-9B 84.7 on the System One Index. GPT-6 Luna Decisions is more accurate on 4 of 16 text datasets, Dexy-Decision-9B on 12. Same cases, same request, same scorer.
| GPT-6 Luna Decisions | Dexy-Decision-9B | |
|---|---|---|
| System One Index | 76.0 | 84.7 |
| English | 75.2 | 82.3 |
| Arabic | 78.8 | 86.8 |
| Vision Index | 81.6 | – |
| Calibration error | 9.8% | 3.5% |
| Median latency | 215 ms | – |
| Cost per 1,000 calls | $0.040 | – |
| GPT-6 Luna Decisions | Dexy-Decision-9B | |
|---|---|---|
| AG News English | 87.5 | 91.5 |
| SST-2 English | 93.5 | 95.5 |
| SST-5 English | 61.5 | 60.0 |
| DAIR Emotion English | 58.5 | 54.0 |
| Banking77 English | 77.5 | 93.0 |
| BoolQ English | 85.5 | 89.5 |
| XNLI English English | 66.7 | 94.7 |
| Typed decisions English | 71.0 | 80.0 |
| XNLI French French | 62.0 | 89.3 |
| XNLI Arabic Arabic | 54.0 | 89.3 |
| MASSIVE scenario Arabic | 66.0 | 89.5 |
| MASSIVE intent Arabic | 84.5 | 91.5 |
| OSACT4 offensive Arabic | 93.5 | 91.5 |
| Tweet sentiment Arabic | 80.0 | 67.0 |
| MIRACL rerank Arabic | 86.0 | 88.0 |
| MIRACL pairwise Arabic | 87.5 | 91.0 |
| GPT-6 Luna Decisions | Dexy-Decision-9B | |
|---|---|---|
| SST-2, as an image | 88.5 | – |
| Arabic sentiment, as an image | 67.0 | – |
| Oxford-IIIT Pet | 87.0 | – |
| VQAv2 yes/no | 80.5 | – |
| CountBench | 85.0 | – |
Model pages: GPT-6 Luna Decisions · Dexy-Decision-9B. Latency for API models is the median per call through OpenRouter, timed from a cloud server with 8 calls in flight.