Part one of three. This release is big enough that we're telling it in three parts.

The short version

We rebuilt how Tresor works through a request — the largest change we've made to the product yet. You'll notice it most on the questions that used to take three prompts and a follow-up.

It decides how hard to think

There is no Instant or Thinking switch any more. Tresor reads the request, then sets its own pace: more reasoning where the problem is genuinely hard, less on the routine steps in between. You choose the model; Tresor works out the depth. Hard questions get the effort they deserve, and simple ones stop waiting behind a long think.

Deep research, on request

Ask for a proper investigation and Tresor shows you the searches and pages it wants to read before anything leaves the sealed environment. Approve them — one by one, or all for this analysis — and it can keep going for up to 20 minutes, following the trail, cross-checking sources, and telling you what it covered. If you'd rather have the answer now, Finish now summarises what it has.

Code, when code is the right tool

For calculations, large tables and anything that needs a real interpreter, Tresor can now run Python. The sandbox has no network, no credentials, and no access to anything you didn't give it. Files it produces come back to you as downloads.

Leaving doesn't stop it

Switch to another chat, reload the page, or close the browser entirely — the run keeps its place. Come back and the progress, the pending approvals and the Stop button are where you left them. No duplicate work, no lost thread.

Until next time

This is the foundation for what comes next: following the same run from another device or network is built and on its way, along with the improvements it makes possible.

Part two covers what this does for documents. Part three is about the app itself.