BizBench - FormulaEval: leaderboard
FormulaEval, a BizBench task (Koncel-Kedziorski et al., 2023): 50 Python functions for financial and accounting formulas, some as methods of balance-sheet, income-statement and cash-flow classes, to complete from a signature and docstring. A completion is correct when its outputs match the reference on 100 random inputs; zero-shot accuracy in percent (higher is better).
Metric: Accuracy (%; 50 financial-formula functions to complete from a signature and docstring, zero-shot, correct when outputs match the reference on 100 random inputs). Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 | 100 | #996 |
| 2 | Mixtral 8x7B | 94 | #4051 |
| 3 | Llama 2 70B | 80 | #4021 |
| 4 | GPT-3.5 | 76 | #2430 |
| 5 | Mistral 7B | 66 | #7784 |
| 6 | mpt-30B | 60 | #8820 |
| 7 | Llama 2 7B | 52 | #9031 |
| 8 | Llama 2 13B | 52 | #8311 |
| 9 | mpt-7B | 48 | #9926 |
| 10 | starcoder | 18 | #9634 |
| 11 | falcon-7B | 14 | #9713 |
| 12 | falcon-40B | 6 | #8304 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=bizbench-formulaeval · How It Works · Data refreshed daily, snapshot 2026-11-12.