Do more capable models fail more often on long agent tasks?

In one study, yes: the two strongest of ten open-weight models led on long-task completion and also had the highest meltdown rates. Across a 23,392-episode study spanning four task lengths, DeepSeek V3 and MiniMax M2.5 scored highest on the paper’s partial-credit measure for very-long tasks (0.87 and 0.89) and melted down on 19% and 13% of very-long episodes, against 0 to 4% for every weaker model tested. A meltdown is a spike in the entropy of the model’s tool calls: it starts repeating a tool with minor argument tweaks or chasing a dead-end subtask instead of making coherent progress.

Weaker models melt down less because they never attempt the ambitious multi-step plans that could go wrong; a shallower strategy completes less of the task but also derails less. Reliability and raw completion rate measure different things, and this is a case where the model that finishes the most is also the one most likely to spiral on its worst runs. The study tested open-weight models specifically; whether proprietary frontier models show the same pattern hasn’t been tested here.