Deep research
A longer, structured version of research mode for questions where you want a report rather than an answer.
How a run goes
-
[1]
Plan
A strong model turns your question into the pieces the run works from.
- Sub-questions.
- A set of facets to investigate.
- The criteria that would mean the question has actually been answered.
-
[2]
Gather, in parallel
One agent per facet
Facet
Facet
Facet
and so on
Each agent finds sources, fetches them, and pulls out specific claims with the quote each one rests on.
-
[3]
Check the citations resolve
Every
[n]in the report must point at a source that was actually fetched, and every key finding must carry at least one.- [1] a source the run fetched holds
- [2] a source the run fetched holds
- [3] nothing that was fetched fails
- [ ] a key finding carrying no citation at all fails
Either failure sends the report back. It is regenerated once.
-
[4]
Decide whether to go again
If the success criteria are not met, it plans another round. It stops when one of these is true:
- The criteria are met
- It hits the budget
- It hits the round cap
The dashed return is the round that goes again: it starts back at the plan.
What comes back
Citations you can follow.
A record of what was dropped along the way.
It appears under Jobs like other long-running work, and survives the browser being closed.
What it costs
Several models ×many fetches ×multiple rounds
This is one of the more expensive things you can run. Set a cap when you start it if the question is open-ended.
Not the same guarantee as research mode
Research mode
- Checks
- That a quote appears verbatim in the page it cites.
- Strength
- The stronger check.
- Reach for it when
- A single claim has to be bulletproof.
Deep research this page
- Checks
- That every citation resolves to a source it really fetched, and that nothing is left uncited.
- Strength
- A weaker guarantee, on far more material.
- Reach for it when
- You need breadth and will read the sources yourself.