Snorkel Finance Reasoning: leaderboard
Snorkel AI's archived Finance Reasoning benchmark: analyst questions needing about 12 steps of reasoning and successive SQL calls over tables converted from 10-K filings, co-created with Snorkel's financial expert network. Claude 3.7 Sonnet scores answers against ground truth, failed traces included; accuracy %, higher is better. Frozen roster.
Metric: Accuracy (%, LLM judge against ground truth). Source: snorkel.ai. Saturation forecast: Around February 2027. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4 | 53.1 | #189 |
| 2 | GPT-5.4 | 52 | #83 |
| 3 | Claude 3.7 Sonnet | 51.89 | #259 |
| 4 | GPT-5 | 51 | #105 |
| 5 | Claude Sonnet 4 | 49.37 | #207 |
| 6 | Claude Opus 4 | 48.1 | #170 |
| 7 | Gemini 3 Pro | 46.84 | #82 |
| 8 | GPT-5 Mini | 46.8 | #193 |
| 9 | O4 Mini | 45.57 | #185 |
| 10 | Claude Opus 4.1 | 45.56 | #146 |
| 11 | GPT-4.1 | 44.3 | #258 |
| 12 | O3 | 43.04 | #132 |
| 13 | Grok 3 | 41.8 | #313 |
| 14 | Grok 4 Fast (Reasoning) | 40.51 | #266 (Grok 4 Fast) |
| 15 | Kimi K2 (Thinking) | 35 | #252 (Kimi K2) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=snorkel-finance-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-05.