Models & Auto mode

Model choice isn't a global preference here. It's a per-thread decision you can change between turns, plus a separate decision about what the sub-agents inside that thread run on.

The model picker

Every thread has a model picker in the composer. It sets which model the lead agent runs on, and you can change it at any point — the next turn simply uses the new one, with the same conversation and the same sandbox behind it. So a thread can start on a fast model while you scope the work, move to a stronger one for the hard middle, and drop back down for cleanup.

The menu leads with Optimize for: Auto, then Intelligence, Balance, and Cost. Turn on Advanced to pick an exact model instead: a grid lists each model with its effort and its input and output price per million tokens. A model's effort submenu sets how hard it thinks, and supported models offer Pro mode — more model work, slower, more tokens. Switching model or effort can ask first when it would discard a warm prompt cache.

The selectable list is a registry maintained by the platform rather than something you configure. New models appear when they're supported; retired ones rotate out, so you're never pinned to a model that no longer exists.

Auto mode

Auto hands the choice to the router instead, choosing among the Intelligence, Balance, and Cost tiers. It picks a tier for each turn rather than for the whole thread; it does not escalate to a stronger tier mid-turn.

The point is that most turns don't need the frontier model. Renaming a symbol, reading a file, answering "where does this live?" — those are cheap. Untangling a race condition isn't. Auto keeps the cheap turns cheap without making you predict, turn by turn, which kind you're about to send.

A provider rate limit doesn't fail the turn either. The turn parks, retries, and after repeated limits is re-routed off that provider — you see a paused turn resume, not an error to re-send.

Sub-agents have their own setting

The lead's model doesn't have to be the swarm's model. Sub-agents get their own per-thread Subagent model setting, in the Options section of the side panel's Extras tab:

This is worth tuning. A coder implementing a scoped slice within the lead's brief, or an explorer reading files, can often run a tier below the lead that's holding the whole plan in context — that's the delegation pattern paying off twice. Some specialized fixer roles keep their own pinned defaults regardless, because the job is narrow and the right model for it is already known.

Model choice is a cost lever

Pricing differs per model, and usage is metered to the billing account selected by the project's Billing setting — see usage & limits for where that shows up. Between Auto, a cheaper tier for sub-agents, and delegation keeping the lead's context small, you have three independent dials on what a thread costs.