How routing decides

Routing picks the right model per question. The rules are written to be read - this page is the whole story. In the app, Settings → Routing shows your models as routing sees them and a live list of recent decisions, and links back here.

You do not need to change anything. An Offline Only AI picks the best model that runs well on this machine. An Online and Offline AI answers with a frontier online model by default and keeps a question on your device when a model here is better for it. The rest of this section is for when you want to know how, or want to lean it one way.

On your device, routing picks by real numbers. As you use each model, the app measures how fast it actually runs on this computer, and ranks by that - not by a spec sheet. Online models are chosen by job instead - one for everyday questions, one for hard ones, one for the live web - since their speed depends on the provider, not on your machine.

The rules

For an Online and Offline AI, each question walks this ladder from the top. The first rule that applies decides.

1. Health stays home. A health question goes to the model you chose for health on your device - never online, whatever the dial says. Online for health is only ever an explicit, asked-every-time choice. See Health questions.
2. Too long for this device goes online. A conversation, document or image that no model here can hold - even after the app makes more room - goes to an online model that can. An attachment only reaches this rule with your consent; see Attachments and images.
3. A short follow-up stays put. "Why?", "shorter please", "and the second one?" - a follow-up that carries no signal of its own stays with the model that answered last, unless it has a live-web cue of its own or the topic has moved on.
4. Current information goes to the live web. A question that needs fresh facts goes to the web-search model, with cited sources and live progress. A clear cue ("today", "latest", "weather") reaches it on every dial; a cue-free question ("is the bridge still closed", "what is the newest model") is caught by meaning. A greeting with "today" in it stays home.
5. Hard questions go to a stronger model. A question the helper grades as hard goes to your Hard online model - except under Local-first, where hard questions stay home.
6. Everything else goes to the Everyday model, unless your device is better. An ordinary question goes to the Everyday online model (GPT-5.6 Luna by default). It stays on your device when the model that would answer here runs at full speed, the question is known easy, and that model scores better than the Everyday model on the task under Frontier-first, or at least as well under Balanced. Local-first never sends an ordinary question out.

Under Local-first only rules 1 to 4 apply: online is for the live web and for what this device cannot hold. The dial is explained in Routing settings.

On your device

Among installed models, routing prefers what runs well on this machine - see Offline models for the fit grades. A model that fits fully on your graphics card runs well; so does a mixture-of-experts model running GPU + RAM. A dense model that does not fit the card is too large here, and routing never picks it. Once you have used a model, its real speed on this computer is known, and "Prefer fastest" ranks by it. A model that takes long to load must be clearly better before routing reloads to it for one question. Projects use tool-capable models only; see Working on projects.

Thinking before answering

A hard question on a model whose reasoning can be switched thinks before it answers; a known-easy question answers directly. The recent decisions in Settings show which ("thinking on", "direct answer").

The router learns, a little

What you do with an answer - Redo on your device, Try this answer online, regenerate - is kept on this computer. Once there is enough of it to mean something, it nudges the as-good rule by at most one point in either direction, and a model here whose answers you keep sending online or regenerating loses a point on that task. Never for health questions. Reset what routing has learned on this computer, in Settings → Routing, forgets it all.

The helper does the grading

The task and difficulty of a question come from the optional helper, and the live-web and health checks by meaning need its small memory model. Without the memory model the health check cannot run, so routing keeps ordinary and hard questions on your device rather than send an unchecked one out - the reply says so. A clear live-web cue still goes to the search model. Without the helper model, difficulty is unknown: an ordinary question goes to the Everyday model, the safe direction for quality.

The reply always says why

The Model button under every reply names the model that answered, where it ran (on device, online, your server), and the routing reason - with one-click second opinions: Redo on your device or Try this answer online. Never a silent substitution. The decision is also written into the conversation's record.

Switching mid-conversation

The model line on an AI's card and the chip in the chat's ask row are switchers: change the model for the next turn without leaving the conversation. The reply still says which model actually answered.

Where the levers are

Every routing setting lives in one place - see Routing settings - and the modes themselves are explained in Routing modes.