Grok 4.5 Tops the Tencent Mixnet Agent Endurance Leaderboard, Completing Over 60% of Tasks Without a Model

According to Dynamic Insight Beating monitoring, the Tencent MetaUnity team, in collaboration with multiple universities, released LHTB to specifically test whether an AI Agent can continuously complete complex tasks within a terminal.

The benchmark consists of 46 tasks covering areas such as software engineering, experiment replication, scientific computing, and multimodal analysis. Each task runs for a maximum of 90 minutes and typically requires hundreds of consecutive steps.

LHTB does not only consider the final success or failure. Partial task completion is also scored based on actual progress. A hidden validator checks files, test results, and program output; self-proclaimed completion by the Agent is not counted.

The initial version of the paper tested 15 models. GPT-5.5 completed 7 tasks, ranking first. Each model performs one task, consuming an average of about 9.9 million tokens and taking around 85 minutes.

The current public results have expanded to 20 models. Grok 4.5 leads with an average score of 0.505, completing 13 tasks. Out of the 46 tasks, there are still 29 tasks that no model has met the completion criteria for.

Most Agents can complete some steps but often run out of time after 90 minutes or stop early before the task is finished, failing to truly complete the complex tasks.