benchgap
Calibration

SWE-bench Verified → AA-SciCode

AA-SciCode is estimated from SWE-bench Verified with a Michaelis–Menten curve fitted on 34 models measured on both: y = 2.0000·x / (2.51037 + x), R² = 0.75, cross-validated error 3.9 pp. It is used for 41 estimates.

Estimated modelSWE-bench VerifiedAA-SciCodeSource
BTL-478.4%47.6%estimated ± 3.9 pp, high confidence
MiMo-V2-Pro78.0%47.4%estimated ± 3.9 pp, high confidence
Claude Sonnet 4.577.2%47.0%estimated ± 3.9 pp, high confidence
MiMo-V2-Omni74.8%45.9%estimated ± 3.9 pp, high confidence
Claude 4.1 Opus74.5%45.8%estimated ± 3.9 pp, high confidence
Step 3.5 Flash74.4%45.7%estimated ± 3.9 pp, high confidence
Flint74.2%45.6%estimated ± 3.9 pp, high confidence
Claude Haiku 4.573.3%45.2%estimated ± 3.9 pp, high confidence
Jais-3-MoVA-36B-A4B72.8%45.0%estimated ± 3.9 pp, high confidence
Claude 4 Sonnet72.7%44.9%estimated ± 3.9 pp, high confidence
MAI-Code-1.1-Flash72.6%44.9%estimated ± 3.9 pp, high confidence
Claude 4 Opus72.5%44.8%estimated ± 3.9 pp, high confidence
Jais-3-32B72.4%44.8%estimated ± 3.9 pp, high confidence
Qwen3.5-27B72.4%44.8%estimated ± 3.9 pp, high confidence
Grok Code Fast 170.8%44.0%estimated ± 3.9 pp, high confidence
Solar Open 270.4%43.8%estimated ± 3.9 pp, high confidence
Jais-3-375B-A23B70.2%43.7%estimated ± 3.9 pp, high confidence
Jais-3-7B70.2%43.7%estimated ± 3.9 pp, high confidence
Qwen3-Coder 480B-A35B69.6%43.4%estimated ± 3.9 pp, high confidence
Qwen3.5-35B-A3B69.2%43.2%estimated ± 3.9 pp, high confidence
DeepSeek V3.166.0%41.6%estimated ± 3.9 pp, high confidence
Kimi K265.8%41.5%estimated ± 3.9 pp, high confidence
Claude 3.7 Sonnet62.3%39.8%estimated ± 3.9 pp, high confidence
Ternary Bonsai 2 27B60.8%39.0%estimated ± 3.9 pp, high confidence
Nemotron 3 Super 100B60.5%38.8%estimated ± 3.9 pp, high confidence
GLM-4.7-Flash59.2%38.2%estimated ± 3.9 pp, high confidence
DeepSeek-R1-052857.6%37.3%estimated ± 3.9 pp, high confidence
MiniMax M1 80k56.0%36.5%estimated ± 3.9 pp, high confidence
GPT-4.154.6%35.7%estimated ± 3.9 pp, high confidence
ZAYA1-74B-Preview53.2%35.0%estimated ± 3.9 pp, high confidence
o3-mini49.3%32.8%estimated ± 3.9 pp, high confidence
DeepSeek-R149.2%32.8%estimated ± 3.9 pp, high confidence
Claude 3.5 Sonnet49.0%32.7%estimated ± 3.9 pp, high confidence
Devstral Small46.8%31.4%estimated ± 3.9 pp, high confidence
Sarvam 105B45.0%30.4%estimated ± 3.9 pp, high confidence
Claude 3.5 Haiku40.6%27.8%estimated ± 3.9 pp, medium confidence
Sarvam 30B34.0%23.9%estimated ± 3.9 pp, medium confidence
Gemini 2.5 Flash-Lite31.6%22.4%estimated ± 3.9 pp, medium confidence
Gemini 2.0 Flash21.4%15.7%estimated ± 3.9 pp, medium confidence
Gemini 2.0 Flash-Lite12.5%9.5%estimated ± 3.9 pp, medium confidence
Gemini 1.5 Flash9.6%7.4%estimated ± 3.9 pp, medium confidence

Fitted on

ModelSWE-bench VerifiedAA-SciCode
Apodex 1.177.7%45.5%
Claude Fable 595.0%61.0%
Claude Opus 4.888.6%54.4%
Claude Opus 596.0%56.4%
Claude Sonnet 585.2%54.3%
DeepSeek V342.0%35.8%
DeepSeek V4 Flash 073179.0%50.3%
DeepSeek V4 Pro 081380.6%51.0%
Gemini 2.5 Pro63.8%46.3%
GPT-OSS 120B62.4%34.0%
GPT-OSS 20B60.7%38.9%
Granite 4.2 30B57.0%37.8%
Granite 4.2 8B47.7%31.5%
Hy3 Preview74.4%48.6%
Inkling77.6%47.0%
Inkling-Small80.2%49.7%
K-EXAONE 2.068.2%42.0%
Kimi K2.680.2%51.5%
MiniCPM5-2B46.4%26.3%
MiniMax M380.5%47.1%
Mistral Medium 3.5 128B77.6%40.2%
Muse Glimmer 30B76.0%44.9%
Nemotron 3.5 Lightning 30B A3B NVFP452.8%32.1%
Nemotron 3.5 Super VL 120B A12B71.1%47.8%
Nemotron 3.5 Super VL 120B A12B NVFP471.1%47.8%
Nemotron 3 Ultra71.9%40.3%
Qwen3.5-122B-A10B72.0%39.7%
Qwen3.5 397B76.2%44.8%
Qwen3.6-27B77.2%42.8%
Qwen3.6-35B-A3B73.4%36.6%
Qwen3.7 Max80.4%49.5%
Qwen3.7 Plus77.7%46.1%
Beam 1.0 501B-A23B81.6%48.5%
Solar Pro 470.6%48.0%