← All benchmarks
Finance· by Column Tax

TaxCalcBench

Can an AI fill out a real US tax return correctly? TaxCalcBench gives the model a taxpayer’s W-2s, 1099s and details, and asks it to produce the complete Form 1040. Every line it computes is checked against the correct return, and the score is the share of lines it gets right.

Benchmark on GitHub ↗ Read the paper ↗
harness (tools)raw model95% CI
By-line accuracy, %
0
20
40
60
80
100
91.4
91.2
90.1
72.8
70.7
gpt-5.5codex
opusclaude-code
sonnetclaude-code
haikuclaude-code
gpt-5.1direct

Same 51 held-out cases for every bar.