Choose a model
Every task runs on a model. You can let Context pick one for you with a routing preset, or pick an exact model yourself. This page explains what each choice does so you can pick well. To send a task, see Send a task.
The model chip
The model chip sits in the composer’s toolbar. Click it to open the model menu.
- Presets come first. Auto, Smart, Balanced and Cheap head the list, in that order. Each one picks a model for you.
- Every other model follows. Each model shows its maker’s mark and name. A check marks the choice in force. Which models appear depends on what your install serves (see Why your list may differ).
- Search filters the list. Type in Search models to find a preset or a model by name.
- Hover for details. On a preset, the details show which model each tier runs and what decides the tier. On a model, the details show its reasoning levels and how it is billed.
On a phone, the same list opens as a sheet from the bottom of the screen.
If the chip is greyed out, no model is available to you here. Hover it to see why, or ask an admin.
Presets and tiers
Context sorts work into four tiers, from quickest to hardest. Your admin picks a model for each tier, so the exact models differ between organizations.
| Tier | Meant for |
|---|---|
| Fast | Quick, simple asks: a fact, a short rewrite, a greeting |
| Balanced | Everyday work: drafting, summarizing, explaining, routine code |
| Deep | Hard work: long multi-step tasks, debugging, design |
| Deepest | The hardest: careful reasoning, proofs, decisions under tradeoffs |
These are the built-in tier names. An admin can rename them, so your menu may show different names.
All four presets read the same tiers. Each preset limits which tiers a task can land on.
| Preset | Tiers it can use | What it does |
|---|---|---|
| Auto | Fast, Balanced, Deep, Deepest | Picks a model for each new task from its first message. |
| Smart | Deep, Deepest | Like Auto, but never below Deep. A quick ask still gets a top model. |
| Balanced | Fast, Balanced | Like Auto, but never above Balanced. No task reaches the top models. |
| Cheap | Fast | Every task runs on the fastest, lowest-cost model. |
When a task’s first message fits a tier outside the preset’s range, the task runs on the nearest tier inside the range. For example, a hard request sent with Balanced runs on the Balanced tier.
How a preset decides
A preset decides once, on the task’s first message.
- The first message is read. When the install has a small model, that model reads the message and picks a tier. If it has none, the message is scored from its wording instead, with no model call. Cheap always uses the wording score, because it has only one tier.
- The task gets that tier’s model. If the small model doesn’t answer, the task goes to the fallback tier your admin set.
- The model stays for the task. Later messages in the same task keep the model the first message got. A task left idle for more than 30 days is decided again on its next message.
To see what a preset chose, look at the footer under a reply. It reads Auto followed by the model’s name, for example “Auto · Model name”. Hover it to see the tier, how the tier was picked, and when it was decided. The model’s name can take a few seconds to appear.
Put the hard part of your request in the first message. A preset reads only that message, so a short “hi” followed by a long brief can land the whole task on a quick tier.
Pick an exact model
Pick a model from the list when you need that model specifically, for example to compare answers or to match what your team tested.
Some models offer Reasoning levels. Hover the model and pick a level, or leave it on Auto to use the model’s default. If a model shows “Reasoning is fixed for this model.”, it has no levels to choose.
Choose for common jobs
| Job | Good choice | Why |
|---|---|---|
| Most everyday tasks | Auto | It matches the model to each task, so quick asks stay quick and hard ones get a stronger model. |
| A quick fact, rewrite or reply | Cheap | The fastest, lowest-cost model handles short asks well. |
| Drafting, summaries and routine code at steady cost | Balanced | Work stays on the two lower tiers, so cost stays predictable. |
| Debugging, design, long multi-step work | Smart | Every task starts on a top tier, even when the first message is short. |
| Careful reasoning or a decision with tradeoffs | Smart, or an exact model with a higher Reasoning level | You get a top model every time, not only when the first message looks hard. |
| Comparing two models on the same brief | An exact model | Each task runs on the model you named, with no routing. |
Cost and credits
On a deployment that bills in credits, the model calls a task makes are charged as credits. Cheap keeps every task on the lowest-cost tier, and Balanced keeps tasks off the top tiers, so both are good picks for light work.
- Estimate before you send. For a video, the composer shows an estimate beside the model chip, such as “About 1,450 credits”. Other task types show no estimate.
- Billing line. A model’s details can show how it is billed, for example “Billed per token, direct to” its maker.
- When credits run out. A running task pauses, and the task page says why, for example “Agent paused: this workspace is out of credits.” or that it paused at the workspace’s spending limit. Click Upgrade to add credits or change plan. Depending on the run, it then picks up where it stopped, or you send your message again. If your plan can’t be changed from the app, the notice reads “This run is paused until your plan has credits.”
- Where usage shows. Settings > Usage shows spend from the billing ledger, with a Top models list and an activity table that names the model behind each chat. Owners on a billing deployment also see Billing, with the plan and credit balance. If you don’t see these, ask an admin.
An agent’s default model
Each agent has a default model that its new tasks start on. You set it in the agent’s setup, in the Model section: “Pick a routing preset or an exact model.”
- Default puts the agent back on the default model. The row names the model it lands on now, for example “Default (Model name)”.
- A preset or an exact model saved here becomes the agent’s default for new tasks. You can still change the model for any one task in the composer.
- If the saved model is no longer offered, the section warns that new tasks run on the default instead, with a Use default button.
If you only view an agent, the section shows its model’s name, or “The deployment default”.
For when to change an agent’s model, see Create an agent.
Models in applets
Applets that use AI choose their model in their own code, from the models your install offers to applets. You don’t pick a model when you use an applet. An admin can limit which models applets may use.
Why your list may differ
The model menu shows only what your install serves and your admins allow.
- The install decides which models exist. Models come from the install’s model gateway, which the install’s operators set up.
- Admins decide where presets route. Organization admins and owners set the model for each tier, so Auto can run different models in two organizations.
- Presets can be off. A preset appears only while the install serves it.
- Images need image-ready tiers. A preset accepts images only when every tier model can read them.
- Models can be retired. If a task’s model is no longer served, a line in the task says so, for example “Model changed: … is no longer available here. This chat now uses …, the default.” If the same model was renamed, the line says “Model updated” and the task continues on it.
Next steps
- Send a task: write a brief, choose an agent and model, and send it.
- Create an agent: set an agent’s default model in its setup.
- Evals: compare models on your own work before you pick one.
- Models and routing: how admins set the tiers, the fallback and the models on offer.