HorizonBench: leaderboard

HorizonBench (2026) tests long-horizon personalization: models read six months of synthetic user conversations in which life events change preferences, then pick the response that fits the current preference among five options. Accuracy in percent over all items; higher is better.

Metric: Accuracy (%) on all 4,245 items (360 users, three generator models) of HorizonBench (5-option questions on which assistant response fits the user's current preference, given six months of synthetic conversation history, about 163K tokens and 4,300 turns; pre-evolution values as hard distractors; items kept only if five validation models all failed them without the history; chance 20); higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 21 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Opus 4.552.8#160
2Claude 3.7 Sonnet47.8#470
3Claude Opus 445.1#326
4Gemini 3.1 Pro (Preview)41.2#99
5Gemini 3 Pro40.9#145
6Claude Sonnet 4.534#261
7Claude 3.5 Haiku32.5#1580
8Claude Sonnet 432.2#410
9Claude 3.5 Sonnet (20240620)29.8#692
10Gemini 2.0 Flash24.7#723
11GPT-4o24.6#751
12Gemini 2.5 Pro22.8#276
13Gemini 2.0 Flash Lite20.9#1050
14Gemini 2.5 Flash Lite19.5#964
15O316.5#237

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=horizonbench · How It Works · Data refreshed daily, snapshot 2026-11-21.