Depth is the chip beside the composer. It sets two dials for the discussion in front of you — how many seats take part, and how many rounds they get — so the cost of a turn is something you choose rather than something you find out afterwards.
- Quick — one model, one pass. For a question you would not ask twice.
- Compare — two models, one round. A second opinion.
- Consensus — the whole room, reviewing until they agree or run out.
- Deep — the whole room, twice as long to disagree. For work you cannot redo.
Auto
Let Each Question Choose Its Depth lets each message pick its own depth, on the Mac, the web and the iPhone alike. It is on by default and follows your account. A short question one model can settle ("what does HTTP 418 mean?") is answered by the lead alone, in one round, and an Ask Everyone button under the answer puts the same question to the whole room with the usual review. A question that names work worth checking, asks the room, asks for a judgment or carries code gets the whole room as before; so does an @mention, which answers as asked, and a side-by-side discussion, where every column always answers.
On the Mac it also escalates from what actually happened rather than from how the question looked. A one-line question that turns out to make a command fail gets reviewers added mid-run. A guess about difficulty is a guess; a failing command is a fact.
Small questions on a cheaper model
Use a Cheaper Model for Small Questions answers the same kind of short question on the seat's cheaper model: the lead's in a discussion, the AI the router picked in a channel. It is on by default and follows your account. It never applies to a seat you named, to a side-by-side discussion (the columns are there to compare the models you chose), to a question with files attached, to code or real work, or to a question about what is true now, such as the news or a price. A seat running on its own command-line tool also needs Small asks on a CLI's reviewing model. Keep the leader's model while it's cheaper still keeps the lead where it is when re-reading a long discussion on the cheaper model would cost more.
The Summary's Routed row says which model answered and why, for example "Small question → cheaper model: Claude answered on" and the model's name. In a channel, the reply's chip says cheaper model.
Keeping the cost down
- A session ceiling stops a discussion when it has spent what you allowed. Set it in Settings, beside the rounds picker.
- Reviewers run on the cheaper tier of their model by default; the lead moves to its cheaper model for small questions, or when a ceiling is three-quarters spent.
- Switching a seat off is the bluntest and most effective control.
See Plans, credit and what a month costs for what the money side looks like.