TaxCalcBench

Fifty Tax Year 2025 federal and state returns built by tax software analysts. The model reads the taxpayer’s W-2s, 1099s, and other PDFs and must calculate the full return, line by line, with no tax engine. A return counts only when every evaluated line is exactly right; lenient and per-line scores show how close it came.

Last updated · Source leaderboard checked daily

Main Insights

Results

TaxCalcBench results, ranked by score (strict), highest first. Last updated .
RankModelProviderScore (strict)Cost (USD / task)Duration (seconds / task)
1Muse Wonder 1.0Meta83.7%$2.71305
2Claude Fable 5.2Anthropic82.0%$2.62264
3Gemini 4 HydrogenGoogle82.0%$0.73123
4Gemini 4 ArgonGoogle79.6%$0.88391
5GPT-6 SolOpenAI76.0%$0.36137
6GPT-6.1 SolOpenAI74.0%$0.32110
7Gemini 4 CarbonGoogle71.4%$1.05521
8GPT-6 AstraOpenAI70.0%$1.59222
9Claude Opus 5.5Anthropic70.0%$3.54439
10Kimi K3.1 ProMoonshot AI66.0%$0.82250
11GPT-6.1 TerraOpenAI62.0%$0.57317
12Claude Fable 5.1Anthropic58.0%$5.28287
13Hy4Tencent55.1%$0.22528
14Muse EmberMeta54.0%$0.080100
15Grok 4.7xAI48.0%$2.53469
16Gemini 3.8 FlashGoogle46.3%$1.06171
17Kimi K3Moonshot AI42.0%$0.58459
18GPT-6 LunaOpenAI36.0%$0.23422
19Muse Spark 1.3Meta32.0%$0.53155
20Mistral Large 4Mistral6.0%$0.12246
21DeepSeek V4.1 FlashDeepSeek4.0%$0.060217
22Gemini 3.1 ProGoogle2.0%——
23Ultramarine 1 FlashPaper Instruments0.0%$0.030188
24GLM 5.3Z.ai0.0%$0.421,372