benchgap
Calibration

SWE-bench Pro → NL2Repo

NL2Repo is estimated from SWE-bench Pro with a linear curve fitted on 28 models measured on both: y = 1.2179·x + -0.2615, R² = 0.60, cross-validated error 6.1 pp. It is used for 16 estimates.

Estimated modelSWE-bench ProNL2RepoSource
Claude Opus 5.589.9%83.3%estimated ± 6.1 pp, low confidence
Claude Sonnet 5.581.3%72.9%estimated ± 6.1 pp, low confidence
Claude Mythos 580.3%71.7%estimated ± 6.1 pp, low confidence
Ling 3.1 Flash68.7%57.6%estimated ± 6.1 pp, low confidence
Beam 1.0 501B-A23B66.1%54.4%estimated ± 6.1 pp, medium confidence
Beam 1.0 501B-A23B FP865.9%54.1%estimated ± 6.1 pp, medium confidence
Beam 1.0 501B-A23B NVFP465.6%53.7%estimated ± 6.1 pp, medium confidence
Sage-A1.561.2%48.4%estimated ± 6.1 pp, medium confidence
Atria Dawn Preview59.6%46.4%estimated ± 6.1 pp, medium confidence
Laguna S 2.159.4%46.2%estimated ± 6.1 pp, medium confidence
GPT-5.255.6%41.6%estimated ± 6.1 pp, medium confidence
Spark-X2.6-4B54.3%40.0%estimated ± 6.1 pp, medium confidence
Claude Opus 4.653.4%38.9%estimated ± 6.1 pp, medium confidence
Laguna XS 2.147.6%31.8%estimated ± 6.1 pp, medium confidence
LongCat-Flash-Lite-Sparse40.6%23.3%estimated ± 6.1 pp, low confidence
LLaDA2.2-flash30.1%10.5%estimated ± 6.1 pp, low confidence

Fitted on

ModelSWE-bench ProNL2Repo
Claude Opus 4.557.1%43.2%
DeepSeek V4 Flash 073152.6%54.2%
DeepSeek V4 Pro 081355.4%61.5%
dots3-note Preview61.0%49.8%
GLM-555.1%35.9%
GLM-5.158.4%42.7%
GLM-5.262.1%48.9%
Hy4 preview65.7%58.9%
Kimi K2.550.7%32.0%
MiniMax M2.756.2%39.8%
MiniMax M359.0%42.1%
Ornith-1.0-35B50.4%34.6%
Ornith-1.0-397B62.2%48.2%
Ornith-1.0-9B42.9%27.2%
Ornith-1.5-35B-A3B59.6%46.2%
Ornith-1.5-397B65.1%59.5%
Ornith-1.5-9B47.5%32.4%
Qwen3.5 397B50.9%32.2%
Qwen3.6-27B53.5%36.2%
Qwen3.6-35B-A3B49.5%29.4%
Qwen 3.6 Max (preview)57.3%42.9%
Qwen3.6 Plus56.6%37.9%
Qwen3.7 Max60.6%47.2%
Qwen3.7 Plus57.6%41.1%
Qwen3.8-27B61.7%42.3%
Qwen3.8-Flash-Next62.5%48.1%
Qwen3.8 Max67.7%55.9%
Qwen3.8-Omni-Flash63.3%48.9%