LEXam: leaderboard
7,537 questions from 340 University of Zurich law exams in English and German: 2,841 open-ended questions graded by an LLM judge plus 4,696 multiple-choice items with 4 to 32 options; 2025, ICLR 2026.
Metric: Multiple-Choice Accuracy (%). Source: lexam-benchmark.github.io. Status: saturation imminent. 31 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5 | 62.65 |
| 2 | Claude Sonnet 4.5 | 58.01 |
| 3 | Claude 3.7 Sonnet | 57.23 |
| 4 | Gemini 2.5 Pro | 55.72 |
| 5 | GPT-5 Mini | 54.82 |
| 6 | GPT-4.1 | 54.4 |
| 7 | GPT-4o | 53.13 |
| 8 | DeepSeek V3.2 Exp | 53.07 |
| 9 | DeepSeek R1 | 52.41 |
| 10 | Llama 4 Maverick | 49.1 |
| 11 | GPT-4.1 Mini | 48.49 |
| 12 | Qwen 3 235B A22B | 48.19 |
| 13 | QwQ-32B | 47.83 |
| 14 | GPT-OSS-120B | 47.71 |
| 15 | GPT-5 Nano | 47.11 |
Interactive version: theaggregate.ai/benchmark?slug=lexam · How It Works · Data refreshed daily, snapshot 2026-09-05.