Deep Research lets you send one request off to gather, read, and synthesise information from across the web, then watch a full report get written live in a side panel. It now runs as a tool on top of your current thread, so you stay on your own model for the conversation and only switch on deep research for the turns that need it.
How to use it
Open the depth menu in the message composer and choose a research tier. This menu also acts as the on/off switch for the turn.
Pick Fast for a quicker pass, or Max for a deeper, more thorough investigation.
Type your question and send it as normal.
The report streams into the right-hand Artifacts panel. You can watch it search, read, think, and write in real time, with sources and visuals included.
Collaborative planning
For bigger questions, you can turn a single request into a plan-and-approve flow. When enabled, the agent proposes a research plan and stops - this takes only seconds and does no searching yet. You then review the plan in the panel, edit it, and approve it. Once approved, the research runs against the plan you signed off on.
Version history
If you re-run a request, the previous report is kept rather than overwritten. Reports on the same message build up as a version chain, so you can always go back to an earlier run.
Tips
Deep research cards appear inline in your message list. Clicking one opens the Artifacts panel so you can follow along or revisit the finished report.
Max covers more ground than Fast - reach for it when a topic is broad or you need wider source coverage, and stick with Fast when you want a quicker answer.
Because deep research runs per turn, your thread keeps using your chosen model for everything else. You're only in research mode for the messages where you switch it on.