Long-Horizon Task Benchmark Exposes Model Weaknesses
iamfakhrealam · x · 2026-07-13
This post summarizes the results of the Long-Horizon Terminal-Bench, which tested 20 frontier models on 46 terminal tasks lasting up to 90 minutes.
Key takeaways:
- Grok 4.5 ranked first
- The best model only solved 13 out of 46 tasks
- 29 tasks were unsolved by any model
- About 55% of runs scored below 0.25
- Expensive models aren't necessarily better: Grok 4.5 costs $11 per task on average, while Claude Fable 5 costs $73
The author concludes that while models perform well on short tasks, they are still far from being "truly autonomous agents" in long-chain execution, debugging, error recovery, and completing complex work.
Related event: Tencent Hunyuan Releases Long-Horizon Terminal Bench Exposing Model Limits(6 posts)→
More from Models
- Local beats cloud: Qwen 3.8 FN handles private student data as well as frontier models — dr_lm · 2026-10-27
- Aidan Gomez: model's CoT said no tool calls right before making them — aidangomez · 2026-10-27
- Nat Lambert: Cohere's model is narrow RL training on internal infra ill-suited for open weights — natolambert · 2026-10-27
- OpenAI teases DOTS is coming to Europe 'soon' — mark_k · 2026-10-27
- Clifford Pickover challenges AI to decipher a message he encoded in Norse runes — pickover · 2026-10-27
- ChatGPT's crawler sample code strips out all Schema markup — VeryWellVersed · 2026-10-27