Calibration

Vals SWE-bench → AA Coding Index

Name: benchgap: LLM benchmark scores, measured and estimated
Creator: benchgap
License: https://creativecommons.org/licenses/by-nc/4.0/

AA Coding Index is estimated from Vals SWE-bench with a Hill curve fitted on 44 models measured on both: y = 0.2185 + (0.8470 − 0.2185)·x^6.00 / (0.70198^6.00 + x^6.00), R² = 0.74, cross-validated error 5.8 pp. It is used for 5 estimates.

Estimated model	Vals SWE-bench	AA Coding Index	Source
Gemini 3 Flash	75.0%	59.4%	estimated ± 5.8 pp, medium confidence
GPT-5.2-Codex	72.4%	56.2%	estimated ± 5.8 pp, medium confidence
MiMo-V2.5	71.0%	54.3%	estimated ± 5.8 pp, medium confidence
Qwen3.5 Flash	64.4%	45.3%	estimated ± 5.8 pp, medium confidence
Gemini 3.1 Flash-Lite	62.8%	43.1%	estimated ± 5.8 pp, medium confidence

Fitted on

Model	Vals SWE-bench	AA Coding Index
Claude Fable 5	95.0%	76.5%
Claude Opus 4.8	88.6%	74.3%
Claude Opus 5	97.0%	78.0%
Claude Sonnet 5	79.6%	71.6%
DeepSeek V4 Flash 0731	88.8%	69.1%
DeepSeek V4 Pro 0813	96.4%	68.8%
Gemini 2.5 Pro	54.4%	33.3%
Gemini 3.1 Pro	78.8%	68.8%
Gemini 3.5 Flash	78.8%	70.1%
Gemini 3.5 Flash-Lite	75.0%	49.3%
Gemini 3.6 Flash	79.6%	69.2%
Gemini 3.7 Flash	80.8%	76.1%
Gemini 3.8 Flash	80.0%	76.3%
GLM-4.7	69.4%	45.3%
GLM-5.1	76.4%	55.8%
GLM-5.2	82.8%	68.8%
GLM-5.3	95.4%	74.8%
GPT-5.4 mini	73.0%	56.1%
GPT-5.4 nano	69.8%	56.1%
GPT-5.5	82.6%	74.9%
GPT-5.6 Luna	93.0%	71.5%
GPT-5.6 Sol	96.2%	77.4%
GPT-5.6 Terra	95.4%	76.7%
Grok 4.3	71.4%	42.3%
Grok 4.5	86.6%	72.5%
Grok 4.6	95.6%	76.8%
Inkling	77.6%	52.1%
Inkling-Small	82.2%	52.9%
Kimi K2.6	76.2%	61.8%
Kimi K2.7 Code	78.2%	60.8%
Kimi K3	93.4%	76.2%
Ling 3.0 Flash	65.2%	50.6%
MiMo-V2.5-Pro	74.0%	60.2%
MiniMax M2.7	73.8%	52.6%
MiniMax M3	75.0%	58.6%
Mistral Medium 3.5 128B	66.4%	46.9%
Muse Spark	74.4%	58.6%
Muse Spark 1.1	82.0%	71.3%
Muse Spark 1.2	86.6%	72.2%
Nemotron 3 Ultra	69.0%	49.3%
Qwen3.6-27B	70.0%	53.7%
Qwen3.6 Plus	73.4%	54.5%
Qwen3.7 Max	68.8%	66.0%
Qwen3.8-27B	86.0%	68.1%