A chat answer is only as good as what you can verify. Deep Research mode exists for the questions where "sounds right" is not good enough.
What it does differently
Instead of answering from model memory, Deep Research:
- Plans — breaks your question into concrete search angles.
- Reads — fetches and reads real sources across the web.
- Cross-checks — compares claims across sources instead of trusting the first hit.
- Cites — the final report links every substantial claim to where it came from.
You see the sources inline and can open any of them. If a claim has no source, the report says so instead of hiding it.
What it costs
Deep Research is a long-running mode — it makes many model calls and web fetches on your behalf. Every one of them is metered and shows up in your billing ledger as part of the run, priced before you commit. No surprise line items.
When to use it
Market overviews, technical comparisons, "what changed since last year" questions, anything you would otherwise spend an afternoon tab-hopping for. Ask it the question you would give a research assistant, then check the receipts.