TaxCalcBench
Benchmark that tests frontier AI models on US tax return calculations.
What it does
Gives AI models a taxpayer's information, as PDFs of W-2s and 1099s for tax year 2025 or JSON for 2024, and scores their Form 1040 return line by line against the expected e-file XML. A Python tool run with uv that needs API keys for the model providers under test; for AI researchers, model providers and tax software developers. Its finding: models cannot calculate tax returns reliably today; MIT.
Where it fits on the map
- Company tax department › Tax Technology, Data & TransformationTax technology & datacovered
Named as a project to start with: An MCP server that gives an agent Brazilian tax lookups, and a benchmark that shows what models get wrong on returns before you trust one.
- Tax authority › Data, IT & AnalyticsData, document AI & AI modelscovered
Named as a project to start with: A benchmark to test a model on returns, and an embedding model tuned to French tax text.
On Learn
- BenchmarkGlossary term
- 04 · Evidence and reviewTrack
Similar projects
Of the same kind, for a shared jurisdiction or tax area. Not a ranking.
- harvey-labs (Legal Agent Benchmark)Legal Agent Benchmark: 2,010 legal tasks with documents and rubrics and a Python harness; 34 tasks are tax.
- LLM-Tax-AbuseCode and US tax-strategy data from the paper on whether LLMs can identify and generate tax abuse.
- PolicyBenchBenchmark of how well frontier AI models estimate household tax and benefit outcomes without tools.
- SARAStatutory reasoning benchmark built on the US Internal Revenue Code.
- TaxBench114 expert-written US tax questions with grading rubrics for testing LLMs; 19 touch transfer pricing or s.482.Inactive since 2025
Maintain this project?
Link to this page from your README, so readers find the facts, the notes and where it fits. The badge is a small SVG served from this site.
[](https://tax-oss.com/projects/taxcalcbench/)