benchgap
Calibration

NL2Repo → AA-SciCode

AA-SciCode is estimated from NL2Repo with a Hill curve fitted on 15 models measured on both: y = 0.2680 + (0.5469 − 0.2680)·x^4.57 / (0.33514^4.57 + x^4.57), R² = 0.78, cross-validated error 3.0 pp. It is used for 14 estimates.

Estimated modelNL2RepoAA-SciCodeSource
Claude Opus 4.543.2%48.0%estimated ± 3.0 pp, high confidence
dots3-note Preview49.8%50.8%estimated ± 3.0 pp, high confidence
Hy4 preview58.9%52.7%estimated ± 3.0 pp, high confidence
Ornith-1.0-35B34.6%41.8%estimated ± 3.0 pp, high confidence
Ornith-1.0-397B48.2%50.2%estimated ± 3.0 pp, high confidence
Ornith-1.0-9B27.2%34.6%estimated ± 3.0 pp, medium confidence
Ornith-1.5-35B-A3B46.2%49.5%estimated ± 3.0 pp, high confidence
Ornith-1.5-397B59.5%52.8%estimated ± 3.0 pp, high confidence
Ornith-1.5-9B32.4%39.7%estimated ± 3.0 pp, high confidence
Qwen 3.6 Max (preview)42.9%47.9%estimated ± 3.0 pp, high confidence
Qwen3.8 Max55.9%52.2%estimated ± 3.0 pp, high confidence
Qwen3.8-Omni-Flash48.9%50.5%estimated ± 3.0 pp, high confidence
Seed 2.1 Pro47.0%49.8%estimated ± 3.0 pp, high confidence
Seed 2.1 Turbo43.7%48.3%estimated ± 3.0 pp, high confidence