Calibration
V* → CountBench
CountBench is estimated from V* with a linear curve fitted on 6 models measured on both: y = 0.2575·x + 0.7284, R² = 0.83, cross-validated error 1.6 pp. It is used for 1 estimate.
| Estimated model | V* | CountBench | Source |
|---|---|---|---|
| Step 3.7 Flash | 95.3% | 97.4% | estimated ± 1.6 pp, medium confidence |
Fitted on
| Model | V* | CountBench |
|---|---|---|
| Claude Opus 4.5 | 67.0% | 90.6% |
| Gemini 3 Pro | 88.0% | 97.3% |
| GPT-5.2 | 75.9% | 91.9% |
| LFM3-VL-2B | 86.4% | 93.0% |
| Qwen3.5 397B | 95.8% | 97.2% |
| Qwen3.6-27B | 94.7% | 97.8% |