HorizonBench: leaderboard
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. 25 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.5 | 52.8 |
| 2 | Claude 3.7 Sonnet | 47.8 |
| 3 | Claude Opus 4 | 45.1 |
| 4 | Gemini 3.1 Pro (Preview) | 41.2 |
| 5 | Gemini 3 Pro | 40.9 |
| 6 | Claude Sonnet 4.5 | 34 |
| 7 | Claude 3.5 Haiku | 32.5 |
| 8 | Claude Sonnet 4 | 32.2 |
| 9 | Claude 3.5 Sonnet (20240620) | 29.8 |
| 10 | Gemini 2.0 Flash | 24.7 |
| 11 | GPT-4o | 24.6 |
| 12 | Gemini 2.5 Pro | 22.8 |
| 13 | Gemini 2.0 Flash Lite | 20.9 |
| 14 | Gemini 2.5 Flash Lite | 19.5 |
| 15 | O3 | 16.5 |
Interactive version: theaggregate.ai/benchmark?slug=horizonbench · How It Works · Data refreshed daily, snapshot 2026-10-07.