Long-Horizon Terminal-Bench: leaderboard
Long-Horizon Terminal-Bench (LHTB) — coding and research agent tasks in a real terminal that unfold over hundreds of steps, graded by dense per-step reward rather than single pass/fail; headline is the mean reward across tasks (0-100).
Metric: Mean Score (%). Source: zli12321.github.io. Saturation forecast: Around March 2027. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4.5 | 50.5 | #48 |
| 2 | Claude Sonnet 5 | 49.7 | #57 |
| 3 | Claude Opus 4.8 | 49.2 | #37 |
| 4 | Claude Fable 5 | 48.7 | #16 |
| 5 | GPT-5.6 Sol | 45.1 | #19 |
| 6 | GPT-5.5 | 44.5 | #31 |
| 7 | MiniMax-M3 | 38.5 | #136 |
| 8 | Claude Sonnet 4.6 | 37.3 | #90 |
| 9 | Kimi K2.7 Code | 36.7 | #89 |
| 10 | GLM-5.2 | 31.6 | #73 |
| 11 | Qwen 3.6 Plus | 31.3 | #120 |
| 12 | DeepSeek V4 Pro | 30.7 | #102 |
| 13 | Qwen 3.7 Max | 29.6 | #75 |
| 14 | Hy3 | 28.8 | #93 |
| 15 | Gemini 3.1 Pro (Preview) | 27.9 | #58 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=long-horizon-terminal-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.