benchgap
Calibration

VITA-Bench → Claw-Eval

Claw-Eval is estimated from VITA-Bench with a Hill curve fitted on 7 models measured on both: y = 0.0000 + (0.6308 − 0.0000)·x^6.00 / (0.16500^6.00 + x^6.00), R² = 0.77, cross-validated error 7.1 pp. It is used for 3 estimates.

Estimated modelVITA-BenchClaw-EvalSource
Sage-A1.545.4%62.9%estimated ± 7.1 pp, low confidence
Agents-A1-4B40.3%62.8%estimated ± 7.1 pp, low confidence
Agents-A138.8%62.7%estimated ± 7.1 pp, low confidence

Fitted on

ModelVITA-BenchClaw-Eval
Claude Opus 4.523.3%59.6%
DeepSeek V3.218.5%40.2%
Qwen3.5 397B43.7%56.8%
Qwen3.6-35B-A3B35.6%68.7%
Qwen3.6 Plus44.3%58.8%
Qwen3.7 Max47.9%65.2%
Qwen3.7 Plus45.6%62.7%