PrinzBench: leaderboard
33 private questions: 25 on a niche area of US law, each answerable from a few statutes, regulations or agency guidance that the model must find online, and 8 asking it to find an obscure piece of public information. Models are accessed mostly through their vendors' chat products, and each question is asked three times. The author grades every reply by hand, not blind, for a score from 0 to 99, higher is better.
Metric: Score (x/99). Source: github.com. Saturation forecast: Around January 2027. 29 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Pro Sol | 91 | #18 |
| 2 | GPT-5.4 (xHigh) | 69 | #79 (GPT-5.4) |
| 3 | Claude Fable 5 (Max) | 56 | #16 (Claude Fable 5) |
| 4 | GPT-5.3 Codex (High) | 52 | #69 (GPT-5.3 Codex) |
| 5 | Gemini 3.1 Pro (Preview) | 50 | #58 |
| 6 | Kimi K3 | 47 | #30 |
| 7 | Grok 4.20 | 43 | #175 |
| 8 | Claude Opus 4.8 (Max) | 42 | #37 (Claude Opus 4.8) |
| 9 | Gemini 3 Flash | 36 | #99 |
| 10 | Gemini 3 Pro | 35 | #80 |
| 11 | Kimi K2.5 (Thinking) | 35 | #144 (Kimi K2.5) |
| 12 | Muse Spark | 31 | #64 |
| 13 | GLM-5.2 | 30 | #73 |
| 14 | Qwen 3 Max | 25 | #208 |
| 15 | Grok 4.1 (Thinking) | 25 | #227 (Grok 4.1) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=prinzbench · How It Works · Data refreshed daily, snapshot 2026-10-20.