benchgap
Calibration

NL2Repo → Vals SWE-bench

Vals SWE-bench is estimated from NL2Repo with a Michaelis–Menten curve fitted on 13 models measured on both: y = 2.0000·x / (0.69040 + x), R² = 0.72, cross-validated error 5.4 pp. It is used for 15 estimates.

Estimated modelNL2RepoVals SWE-benchSource
Claude Opus 4.543.2%77.0%estimated ± 5.4 pp, medium confidence
DeepSeek V4.1 Flash65.4%97.3%estimated ± 5.4 pp, low confidence
dots3-note Preview49.8%83.8%estimated ± 5.4 pp, medium confidence
Hy4 preview58.9%92.1%estimated ± 5.4 pp, medium confidence
Ornith-1.0-35B34.6%66.8%estimated ± 5.4 pp, low confidence
Ornith-1.0-397B48.2%82.2%estimated ± 5.4 pp, medium confidence
Ornith-1.0-9B27.2%56.5%estimated ± 5.4 pp, low confidence
Ornith-1.5-35B-A3B46.2%80.2%estimated ± 5.4 pp, medium confidence
Ornith-1.5-397B59.5%92.6%estimated ± 5.4 pp, medium confidence
Ornith-1.5-9B32.4%63.9%estimated ± 5.4 pp, low confidence
Qwen3.6-35B-A3B29.4%59.7%estimated ± 5.4 pp, low confidence
Qwen3.8-Flash-Next48.1%82.1%estimated ± 5.4 pp, medium confidence
Qwen3.8-Omni-Flash48.9%82.9%estimated ± 5.4 pp, medium confidence
Seed 2.1 Pro47.0%81.0%estimated ± 5.4 pp, medium confidence
Seed 2.1 Turbo43.7%77.5%estimated ± 5.4 pp, medium confidence