HorizonBench - Evolved Preferences: 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 Accuracy in percent on items whose preference changed after a life event, with the outdated preference as a distractor; higher is better.

Metric: Accuracy (%) on the 2,484 items whose target preference changed after a life event 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 4.551.3#165
2Claude 3.7 Sonnet46.7#455
3Claude Opus 444.4#332
4Gemini 3.1 Pro (Preview)38.4#102
5Gemini 3 Pro38.4#146
6Claude Sonnet 4.532.6#267
7Claude 3.5 Haiku31.8#1597
8Claude Sonnet 431.4#414
9Claude 3.5 Sonnet (20240620)29.2#697
10Gemini 2.0 Flash24.2#731
11GPT-4o23.9#754
12Gemini 2.5 Pro21.5#280
13Gemini 2.0 Flash Lite20.3#1064
14Gemini 2.5 Flash Lite18.2#972
15O115.5#407

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=horizonbench-evolved-preferences · How It Works · Data refreshed daily, snapshot 2026-11-22.