Atlas

Benchmarks

← All benchmarks

TaxCalcBench TY25

Professional Work · 2026-06-22

TaxCalcBench TY25 evaluates complete federal and state tax-return calculation from realistic taxpayer PDFs and companion structured data. Its headline strict score is the percentage of 50 cases for which every evaluated line on Form 1040, California Form 540, Illinois Form IL-1040, New York Form IT-201, or Virginia Form 760 exactly matches the expected return.

Top models (higher is better)

ModelScore
Claude Fable 5.140.0
Claude Opus 5.540.0
GPT-6 Astra38.0
GPT-6.1 Sol36.0
GPT-5.6 Sol30.0
Claude Fable 528.0
GPT-6 Sol28.0
GPT-5.526.0
Claude Opus 522.0
Muse Wonder22.0
Opus 4.820.0
Claude Sonnet 5.518.0
GPT-6 Luna14.0
Gemini 3.7 Flash12.0
Gemini 3.6 Flash10.0
Muse Spark 1.28.0
Muse Spark 1.38.0
Sonnet 56.0
Kimi K36.0
DeepSeek V4.1 Flash4.0
Gemini 3.5 Flash4.0
Gemini 3.8 Flash4.0
Claude Haiku 5.52.0
Gemini 3.1 Pro Preview2.0
GLM-5.30.0
Mistral Large 4 Preview0.0
Loading Atlas data…