UnderwritingBench: Bank Statements

Labelling the transactions on small-business bank statements, the first step of cash-flow underwriting. Seventy-two three-month statements carry 25,593 transactions, each labelled with one of 25 categories, and 32.6% of them accept more than one label because an underwriter could defensibly choose either. Grading is a programmatic match against the accepted labels, with no LLM judge.

Last updated

Main Insights

Results

UnderwritingBench: Bank Statements results, ranked by score (accuracy), highest first. Last updated .
RankModelProviderScore (accuracy)Cost (USD / task)Duration (seconds / task)
1Gemini 4 ArgonGoogle95.6%$0.31298
2Claude Opus 5.5Anthropic95.1%$0.89192
3Kimi K3Moonshot AI93.9%$0.74594
4GLM 5.3Z.ai93.9%$0.651,161
5GPT-6 SolOpenAI93.9%$0.26139
6GPT-6.1 SolOpenAI93.8%$0.24134
7GPT-6 AstraOpenAI93.8%$1.17122
8Muse Spark OpenMeta93.4%$0.221,370
9DeepSeek V4.1 Flash (1017)DeepSeek93.2%$0.070309
10pplx-decider v1.1 27BPerplexity92.7%$0.028—
11GPT-5.6 LunaOpenAI92.7%$0.032164
12Mammoth-1 27BBase Labs92.5%—31
13GPT-6 LunaOpenAI92.4%$0.015198
14Mercury DecideInception91.8%$0.001011
15Clef 27BCloudflare91.5%$0.04655