BizBench - CodeFinQA: leaderboard
CodeFinQA, a BizBench task (Koncel-Kedziorski et al., 2023): 844 FinQA questions over earnings-report text and tables, answered by writing a Python program whose output must be within 1% of the reference. Three-shot accuracy in percent (higher is better).
Metric: Accuracy (%; 844 FinQA questions over report text and tables answered by generating Python code, answer within 1% of the reference, 3-shot). Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 | 78.8 | #996 |
| 2 | GPT-3.5 | 67.5 | #2430 |
| 3 | Mixtral 8x7B | 58.5 | #4051 |
| 4 | Llama 2 70B | 57.3 | #4021 |
| 5 | Mistral 7B | 48.8 | #7784 |
| 6 | Llama 2 13B | 33.4 | #8311 |
| 7 | starcoder | 31.2 | #9634 |
| 8 | mpt-30B | 31 | #8820 |
| 9 | Llama 2 7B | 21.9 | #9031 |
| 10 | falcon-40B | 18.4 | #8304 |
| 11 | mpt-7B | 6.6 | #9926 |
| 12 | falcon-7B | 2 | #9713 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=bizbench-codefinqa · How It Works · Data refreshed daily, snapshot 2026-11-12.