Snorkel Finance Reasoning
Professional Work · 2026-06-05
Snorkel Finance Reasoning evaluates tool-using agents on multi-step financial questions grounded in 10-K filings. Claude Sonnet 3.7 grades final-answer correctness and completeness against ground truth, and the leaderboard reports accuracy across all traces, including failed or prematurely terminated runs.
Top models (higher is better)
| Model | Score |
|---|---|
| Grok 4 | 53.1 |
| GPT-5.4 | 52.0 |
| Claude 3.7 Sonnet | 51.9 |
| GPT-5 | 51.0 |
| Sonnet 4 | 49.4 |
| Opus 4 | 48.1 |
| Gemini 3 Pro Preview | 46.8 |
| GPT-5 Mini | 46.8 |
| O4 Mini | 45.6 |
| Opus 4.1 | 45.6 |
| GPT-4.1 | 44.3 |
| o3 | 43.0 |
| Grok 3 | 41.8 |
| Grok 4 Fast | 40.5 |
| Llama 3.3 Nemotron Super 49B V1.5 | 35.4 |
| Kimi K2 Thinking | 35.0 |
| Gemini 2.5 Pro | 34.6 |
| Nova Premier | 34.2 |
| Gemini 2.5 Flash | 32.0 |
| gpt-oss-120b | 31.6 |
| o3-mini | 30.4 |
| GPT-5 Nano | 26.6 |
| Qwen3-235B-A22B | 17.7 |
| Magistral Medium 1.0 | 13.9 |
| Nova Pro | 12.7 |
| Mistral Large 1.0 | 10.1 |