Deterministic open-source tax engine hits 96% on TaxCalcBench, beating GPT-5 Sol and Fable 5
An open-source US tax engine just posted the highest score on TaxCalcBench to date, 96%, beating frontier reasoning models GPT-5 Sol and Claude Fable 5. Its edge is architectural: an LLM handles language but never the math, and every figure is derived from codified statute into a proof tree anyone can verify offline. Where it can't derive an answer, it refuses to guess.