SnorkelUnderwrite: leaderboard
Snorkel AI's archived SnorkelUnderwrite benchmark: multi-turn commercial insurance underwriting conversations across six small-business task types, in which agents query underwriting tables and ask the underwriter clarifying questions. Overall accuracy judged by GPT-4.1 (94.5% agreement with human annotation); %, higher is better. Frozen roster.
Metric: Overall accuracy (%, LLM-as-a-judge). Source: snorkel.ai. Saturation forecast: Estimated already saturated. 33 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 91 | #83 |
| 2 | Claude Opus 4.1 | 86.3 | #146 |
| 3 | GPT-5 | 83.33 | #105 |
| 4 | Grok 4 | 83.3 | #189 |
| 5 | Grok 4 Fast (Reasoning) | 81.33 | #266 (Grok 4 Fast) |
| 6 | Grok 3 | 78 | #313 |
| 7 | O4 Mini | 78 | #185 |
| 8 | O3 | 77 | #132 |
| 9 | Claude Opus 4 | 77 | #170 |
| 10 | Claude 3.7 Sonnet | 74.6 | #259 |
| 11 | Claude Sonnet 4 | 72.3 | #207 |
| 12 | GPT-5 Mini | 71.67 | #193 |
| 13 | Kimi K2 (Thinking) | 71.3 | #252 (Kimi K2) |
| 14 | GPT-4.1 | 70.6 | #258 |
| 15 | Gemini 2.5 Flash | 61 | #255 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=snorkelunderwrite · How It Works · Data refreshed daily, snapshot 2026-11-05.