A deep analysis run fetches filings, pulls price history, reads the relevant disclosures, and only then starts writing. Most of the elapsed time is retrieval and verification, not generation.
That exceeds every reasonable request timeout, so the work runs as a queued background job. Which means "generating" is a real state in our interface, designed on purpose: what is running, what it has already gathered, and what it will look like when it lands.
Waiting is part of the product
The alternative — answering instantly from the model's memory — is available to us and we decline it. An instant answer about a company's filings, produced without reading the filings, is the exact failure mode this product exists to avoid.
We would rather make you wait four minutes than hand you something confident and unsourced in four hundred milliseconds.