WildTrace: leaderboard
Metric: Rubric credit (%; mean of three non-contestant LLM judges (Claude Sonnet 4.6, Qwen3.5-Plus, Gemini 2.5 Flash) grading answers against criterion-level rubrics, over 481 multi-hop questions on 214 long technical reports and literary works given in full without evidence spans, up to about 1M tokens; a document longer than the model context window is not sent and scores zero; higher is better). Source: arxiv.org. Saturation forecast: Around December 2026. 18 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 75.3 |
| 2 | Claude Opus 4.8 | 72.6 |
| 3 | GPT-5.5 | 71 |
| 4 | GLM-5.2 | 70.9 |
| 5 | Qwen 3.7 Max | 70.1 |
| 6 | Claude Opus 4.6 | 69.2 |
| 7 | Qwen 3.6 Plus | 66.7 |
| 8 | Qwen 3.7 Plus | 65.7 |
| 9 | GPT-5.4 | 62.2 |
| 10 | Gemini 2.5 Pro | 61.4 |
| 11 | GPT-4.1 | 60.4 |
| 12 | GPT-5.1 | 42.1 |
| 13 | Qwen 3 Max | 35.2 |
| 14 | Kimi K2.6 | 35.1 |
| 15 | DeepSeek V3.2 | 33 |
Interactive version: theaggregate.ai/benchmark?slug=wildtrace · How It Works · Data refreshed daily, snapshot 2026-09-29.