Four deep research APIs tested on quality and token cost

An internal research tool needed to take a messy question, read many sources, and return an answer that a person could act on, with citations that held up. Four hosted deep research APIs were tested on the same task to decide which one to connect. The task was designed to break shallow AI agents: the first three sources contradicted each other, one source was quietly out of date, and the correct answer required noticing that the question itself contained a false premise.

The scoring checked whether the final answer caught that false premise, whether every claim linked back to a real source, and how many tool calls and tokens were spent. The main difference was not how many pages each system could search. The real gap appeared when sources disagreed.

The weaker two systems followed whichever source they saw last and gave a confident wrong answer, while the better two called out the conflict and resolved it.