What it can do
Search returns six results, or ten when analysis is on. Read a page pulls the readable text, up to about 20,000 characters. If a page comes back nearly empty the assistant is told it was blocked or needs JavaScript, rather than silently reporting nothing. Find on a page pulls the passages matching a question out of a long document, which is how it handles a large PDF without reading all of it. The assistant reads pages freely from sites that came back in a search, and from sites you linked in the chat yourself. Any other site asks you first, with the full address shown, so text planted in a page or a file cannot quietly send your conversation somewhere. Research subagents cannot ask, so they only read sites from their own searches and from links you gave. Research hands the whole question to a subagent that searches, reads and returns an answer with its sources. Shallow research spends about twelve steps, deep about twice that. It runs its searches and page reads side by side rather than one after another. Find photos searches Unsplash and Pexels, then openly licensed photos on Openverse and Wikimedia Commons, and returns only photos that may be used on a business website: free licences, public domain, CC BY and CC BY-SA. Unsplash+ premium photos and non-commercial or no-derivatives licences are left out. With a Computer attached, the assistant puts the photos it picks straight into the project: each one is checked, resized and saved as WebP underpublic/images, and public/images/ATTRIBUTION.md records where it came from.
A photo whose licence asks for credit is also credited on the site itself,
usually in the footer. A photo search counts as one search, and adding photos
counts as one page read, however many photos it adds.
Sources
Any answer built from the web carries a sources button. Opening it lists every page referenced, with its site, title, description and preview image.Daily limits
On paid plans these also draw from credits at €0.006 a
search, €0.003 a page read and €0.05 a deep search.
Results are cached for a day, so asking the same thing twice does not cost
twice.
When a limit is reached the assistant is told to stop searching and report what
it already found, rather than failing mid answer.