XpertBench: leaderboard
Metric: Weighted rubric score (%) on the XpertBench-Gold subset (245 expert-written open-ended professional tasks, 15-40 weighted binary checkpoints each); ShotJudge: Gemini 2.5 Pro judges every checkpoint with an expert-graded GPT-5 response as one-shot exemplar; overall over all 245 tasks; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 14 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.6 (Thinking) | 66.2 |
| 2 | GPT-5.4 (High) | 64.78 |
| 3 | GPT-5.2 (High) | 61.03 |
| 4 | GPT-5 (High) | 58.7 |
| 5 | Gemini 3.1 Pro (Preview) | 55.05 |
| 6 | Gemini 3 Pro | 53.43 |
| 7 | Kimi K2 (Thinking) | 53.3 |
| 8 | Gemini 3 Flash | 50.98 |
| 9 | Claude Sonnet 4.5 (Thinking) | 46 |
| 10 | DeepSeek V3.1 Terminus (Thinking) | 45.1 |
| 11 | GLM-4.6 | 44.3 |
| 12 | Doubao-Seed-1.6 (Thinking) | 42.1 |
| 13 | Gemini 2.5 Pro | 38.2 |
Interactive version: theaggregate.ai/benchmark?slug=xpertbench · How It Works · Data refreshed daily, snapshot 2026-10-07.