The one thing LLMs still can't do

Large Language Models (LLMs) like GPT 6 Astra or Opus 5.5 struggle with delegating tasks to the most appropriate sub-model. When given a coordinating role, they tend to choose a model similar to themselves for implementation, often resulting in overkill or underkill for the task. This inability to correctly delegate, despite their superhuman capabilities in other areas, remains a significant challenge, with no current model demonstrating sufficient proficiency.

This report highlights a specific, persistent limitation of LLMs like GPT 6 Astra and Opus 5.5, unlike other capabilities where they excel.