Calibration
Toolathlon-Verified → OSWorld 2.0
OSWorld 2.0 is estimated from Toolathlon-Verified with a offset logistic curve fitted on 5 models measured on both: y = 0.1939 + (0.7079 − 0.1939) / (1 + exp(−199.98·(x − 0.7780))), R² = 0.99, cross-validated error 5.5 pp. It is used for 10 estimates.
| Estimated model | Toolathlon-Verified | OSWorld 2.0 | Source |
|---|---|---|---|
| GLM-5.3-Flash | 78.4% | 59.0% | estimated ± 5.5 pp, low confidence |
| Claude Sonnet 5.5 | 77.8% | 45.2% | estimated ± 5.5 pp, low confidence |
| DeepSeek V4 Pro 0813 | 74.1% | 19.4% | estimated ± 5.5 pp, low confidence |
| Hy4 preview | 74.1% | 19.4% | estimated ± 5.5 pp, low confidence |
| Step 5 Preview | 74.1% | 19.4% | estimated ± 5.5 pp, low confidence |
| Kimi K3 | 73.2% | 19.4% | estimated ± 5.5 pp, low confidence |
| GLM-5.3 | 73.0% | 19.4% | estimated ± 5.5 pp, low confidence |
| Muse Spark Open | 72.8% | 19.4% | estimated ± 5.5 pp, low confidence |
| Ornith-1.5-397B | 71.2% | 19.4% | estimated ± 5.5 pp, low confidence |
| DeepSeek V4 Flash 0731 | 70.3% | 19.4% | estimated ± 5.5 pp, low confidence |
Fitted on
| Model | Toolathlon-Verified | OSWorld 2.0 |
|---|---|---|
| Claude Fable 5.1 | 77.8% | 41.7% |
| Claude Opus 5 | 80.6% | 70.6% |
| Claude Opus 5.5 | 77.8% | 48.7% |
| Qwen3.8-Flash-Next | 73.5% | 19.4% |
| Qwen3.8 Max | 72.5% | 19.4% |