Horizon: leaderboard
Horizon RLM-harness long-horizon agent learning. Other harnesses and difficulty cohorts have separate boards. Pass rate uses actual non-null graded cases; incomplete source cohorts retain their observed denominator and an explicit warning.
Metric: Task Pass Rate (%). Source: www.orinlabs.ai. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | claude-opus-4.8 | 55.9 | |
| 2 | gpt-5 | 50.77 | |
| 3 | claude-sonnet-4.6 | 49.74 | |
| 4 | claude-haiku-4.5 | 38.46 | |
| 5 | gpt-5-mini | 24.62 |
Interactive version: theaggregate.ai/benchmark?slug=horizon · How It Works · Data refreshed daily, snapshot 2026-10-20.