PredicateLongBench - Clustered Near-Sorted Decoys (LongBench): leaderboard

Metric: Accuracy (%; share of the 93 LongBench v2 documents under 170K tokens used as word lists whose answer is exactly the one word sequence satisfying the predicate, locate the sorted run with 8 near-sorted decoys clustered next to it; up to 16K output tokens including reasoning, an answer past the cap counts as wrong). Source: arxiv.org. Saturation forecast: Around December 2026. 9 models tracked.

Top models

#ModelScore
1Claude Opus 4.6 (Thinking)93
2GPT-5.4 (High)68
3Gemini 3.1 Pro (Preview) (High)56
4Gemini 3.1 Pro (Preview) (Low)50
5Claude Opus 4.635
6GLM-5.112
7GPT-5.4 (Non-reasoning)4
8Qwen 3.5 397B A17B0
9MiniMax-M2.70

Interactive version: theaggregate.ai/benchmark?slug=predicatelongbench-clustered-near-sorted-decoys-longbench · How It Works · Data refreshed daily, snapshot 2026-09-29.