Deep Research with Real Citations

Deep research is the difference between asking a question and commissioning an answer. You give it a real question; it plans, searches, reads, notices what is still missing, goes back for it, and writes a report with numbered citations you can open.

The loop

1. Plan — break the question into sub-questions worth answering separately.
2. Search — run those as web searches, in parallel.
3. Read — fetch the promising pages and extract what is relevant.
4. Follow up — identify the gaps the first round left and search again.
5. Report — write it, with every claim carrying a citation number that resolves to a source you can open.

What the model decides and what the code decides

The model picks the queries, judges the pages and writes the prose. Everything else — citation numbering, source deduplication, token budgets, how many rounds, when to stop — is ordinary code with unit tests. This split is the reason it works on a phone-sized model: a small model asked to also keep track of its own citation numbering will get the numbering wrong, and a report with wrong citations is worse than no report.

It keeps the receipts

Every run is saved with its sources and its findings, so you can come back to a report from last month and still see what it was built on. Results pair naturally with notebooks, where the sources can live alongside your own notes.

Bounded, and stoppable

Hard token budgets, per-agent timeouts, a fan-out width that comes from BatteryGuard, and a Stop button that works at any point. Research is the most expensive thing the app does after image generation, and it is treated that way.

Availability

Behind the one-time $2.99 unlock in Lite, included in the seven-day trial, and unconditional in the $2 app. Ordinary sub-agents and web search are free in both.