Surprise! The new Explore Tool is already here. I was going to post a preview today but some process improvements we've made recently got us shipping ahead of schedule!

Explore is available for all users above the base tier with (minimum) gpt 4.1 mini/nano, Claude 3.5 Haiku, and Grok 3 mini as research workers. Higher tiers will receive the full suite of models including GPT 4.1, Grok 3, Grok 3.5 (if it ever comes out), Claude 4 Sonnet, and Gemini 2.5 pro. Before I get in to the preview I'll note that Grok is a surprisingly capable researcher, even grok 3 mini has very strong result quality and frequently bests 4.1 mini/nano.

Here's what it looks like in the tools menu:

If you have no previous explore jobs running it'll bring you straight to the new job page:

Enter your question, select whether or not you want the AI to "self drive" (automated feedback on the results) or whether you want to submit feedback yourself during the explore job. Feedback is used to "steer" the AI in additional research directions. For example you could start by searching for information about how guitars are made and then submit feedback asking for more information on how to play the guitar and have both synthesized in to a final report.

Before hitting "start research", if you want to make all of the data (digested web pages, thoughts/summaries/drafts and the final report) easily searchable on your node enter a tag to use for the request. All of the results will become digests tagged with what you selected as well as "yourselection-digests", "yourselection-summaries", "yourselection-final". These can then be easily pulled in to the analyze tool by tag if you want "chat with report."

After hitting "start research" you'll be taken to the main explore job page:

As the AI works through the job you'll see thoughts, sources, and drafts in the center stage info pane. There is also a knowledge graph section at the bottom of the screen that can be used to audit search results and (once the system is done browsing the web results for you) digested summaries of what was found:

(All of the links gathered here are clickable, so if you want to do some browsing of your own in tandem with the AI we encourage you to do so!)

The final reports are a little more robust than standard deep research results especially considering we can obtain them using very small models:

You can also instruct the AI to include local data from your node in searches once you've gotten the hang of it - this connects to the existing vector search and ingestion APIs so there's nothing unfamiliar here besides the user interface. If you're struggling to get the AI to query exactly what you want you can figure the query in vector search out yourself and explicitly provide it.

One thing I will note - the job storage is ephemeral, meaning that it refreshes and clears periodically both for security and performance reasons. The results are saved to your node's digests, so, if a job is no longer visible in the explore page you can still easily find it by tag or by providing a relevant natural language query to vector search.