Not surprising.

Their solution to everything LLM has been to add intermediary LLMs to manage the one we interact with to add "guardrails" or point them to a calculator.

It shouldn't be surprising that they utterly fail at basic tasks on their own.