Routing settings
Settings → Routing has two parts. Offline models is for everyone: it sets how both Auto modes pick among the models on this computer. Online models is for AIs set to Auto - Online and Offline, with a plan: when a question goes online, and to which model. Nothing here needs changing, and AIs set to a specific model are never routed. The levers are for leaning it your way.
Offline models
The models on this computer, from the Offline Models page. Both Auto modes - Offline Only and Online and Offline - pick from them.
Choosing a model
What the pick leans toward among the models you have downloaded.
| Lean | What it means |
|---|---|
| Prefer fastest | Measured speed on this computer wins; until measured, a model that fits fully on your graphics card. |
| Balanced (default) | The best capability that still runs well. |
| Prefer strongest | The most capable model, even if slower. |
Routing never picks a model that is too large for this machine, whichever you choose. Pausing a model on the Offline Models page also removes it from the Auto pool.
Health questions
Health questions stay on your device. This chooses which of your downloaded models answers them - a medical model gives more grounded answers when you have one. With a plan you can choose an online model too; each health question then asks before anything leaves your device. See Health questions.
Project work and your server
Shown once Projects is installed. Project work picks differently: only models that can drive tools are ever used, and the model never changes mid-session. On this computer, quality wins: the most capable model that runs here takes it, by capability, then size, then fit, and a weaker model already loaded gives way, one reload per session. The same pick serves an AI's tools in chat.
Project work runs on matters once your own server is connected: Auto, This computer, or Your server. With your server chosen, the session is told the server's reading room, and a server whose room is smaller than the session needs keeps project work here and says why. A turn served by the server says so on the turn.
Online models
Only for AIs set to Auto - Online and Offline, and only offered with a plan - without one this part is a single line saying so. Health questions always stay on your device; Offline Only AIs never go online.
How much goes online
One dial sets where an ordinary question is answered. Questions that need the live web go online; health questions never do.
| Dial | What it means |
|---|---|
| Online-first (default) | Online unless a model here is better for the question. |
| Balanced | Your device answers when it is nearly as good. |
| Local-first | Online only for the live web; hard questions stay home. |
Once you have asked a few questions, the line above the dial says about how many of your recent ones went online. Where your answers came from, under the dial, shows your last questions by where they were answered, and this plan period's online use by kind of model (everyday, stronger, web search) with what it cost. When the everyday model took part, it adds an estimate at list prices of the same questions all on the strongest model, labelled as an estimate.
Online models by job
When a question goes online, these are the models that take it. Each picker names a recommended model; change it if you would rather use another.
| Job | Takes |
|---|---|
| Everyday questions | Most questions - frontier quality at the lowest price. GPT-6 Luna by default. |
| Hard questions | Deep reasoning, tricky code, complex math - one picker for all of them. |
| Needs current information | Searches the live web and cites sources. |
The recommended names come from the online catalog: a model the catalog marks as the default for a job takes that slot, so new models arrive without an app update. The Use for menu on each card of the Online Models page sets the same picks.
Models automatic routing may use
Routing chooses only from these by itself: the model for each kind of question, and a stand-in from another provider when a model's provider is busy or down. The list is grouped by provider, each with where it is based and how many of its models are on; open a provider for a switch per model. Models from DeepSeek, Moonshot AI, Qwen and Z.ai start off. A model new to the catalog follows its provider's switch.
A model you choose by name - for a job above, or for an AI - is used either way. The Auto switch on each Online Models card is the same setting, and a paused model is out of automatic routing too.
When a model is busy, erroring or not answering before its first word, an Online and Offline AI asks a model of the same kind from another provider you allow, once; the reply's model note says which and why. An AI set to one online model is never switched: its error offers the other model beside Retry. A model that is failing for many people right now is left out of automatic picks until it recovers, and its card says so.
Send attachments to online models
Off by default; see Attachments and images.
Project work online
Shown once Projects is installed.
- Agent work on projects - the online model for the agent and for the helpers it hands file changes, builds and tests to; only models that can drive tools are listed.
- Planning - the online model for the helper that plans before changes: it reads the project and writes the plan, and changes nothing. Helpers that only search the project run on this computer when a capable model fits (below).
- Do simple project side-work on this device - on by default: searching and reading fan-outs run on your device when a capable model runs comfortably here (a fast fit, 8+ tokens a second once measured) - free and private. Otherwise they use the session's online model.
- Keep whole project sessions on this device when possible - off by default: a session goes online unless you ask otherwise. On, a session stays fully on your device whenever a capable model runs comfortably here - slower, but free and private. The Model button tells you when a local model would have been as capable. Under Local-first, agent work never goes online at all.
How much a project session may do unasked lives in Settings → Agent and on each project's chip; see Permissions and approvals.
How routing decides
Routing nudges its choices by what you do with answers - Redo on your device, Try this answer online, regenerate - by at most a point, never for health questions. Reset what routing has learned on this computer forgets it. See How routing decides.
A collapsed panel shows the numbers routing ranks on, for every model you have downloaded: how it runs here (Full speed, GPU + RAM with the split in use, Slower, or Too large), its measured speed and load time on this computer, how much it can read at once, its coding, reasoning and math scores, and whether it can drive project work. Speed and load time fill in as you use each model; they are your machine's numbers, never a benchmark's.
Under it, Recent decisions lists the last decisions with their reasons, whether the answer thought first ("thinking on" or "direct answer"), and whether what routing has learned played a part ("adjusted from your feedback").