Choosing a model
Which model backs the agent, and how to match it to the task.
The agent is backed by a model you choose, and you can change it any time from the picker under the chat input. The model shapes how the agent reasons, how well it handles hard problems, and what the work costs. Picking well is mostly about matching the model to the task.
The models
| Model | Good for |
|---|---|
| Auto | The fast default. Free and unmetered on every plan. Routine work like reformatting, simple transforms, "what's the shape of this," and straightforward plots. |
| DeepSeek V4 Pro | A capable mid-tier model with adjustable reasoning. A solid step up from Auto for everyday analysis. |
| Kimi K2.6 | Another strong mid-tier option. |
| Sonnet 4.6 | Claude Sonnet, a frontier model that handles real analysis and modeling well. |
| Opus 4.8 | Claude Opus, the most capable. For genuinely hard problems: subtle debugging, designing an approach, multi-step work where a wrong turn is costly. |
| Opus 4.7 | The previous Opus, also available. |
Auto runs on every plan. The paid models unlock on the paid plans.
How to choose
A rough rule that holds up:
| The task is… | Reach for |
|---|---|
| Mechanical, like a reformat, a rename, or a simple plot | Auto |
| Normal analysis and modeling | DeepSeek V4 Pro or Kimi |
| Hard, like a tricky bug, approach design, or a high-stakes step | Sonnet or Opus |
You're not locked in. Start a session on Auto, switch up when you hit something that needs more, and switch back down after. Because the choice is per-message, you spend the expensive model only where it earns it.
The model is one dial, and reasoning effort is the other. A mid-tier model on high effort can beat a top model on low effort for some tasks, and cost less while doing it. Both are worth understanding.
Defaults and custom instructions
Set a preferred model and effort in Settings → Models so new sessions start where you like, and per-message overrides still apply on top. The same screen has custom instructions: standing guidance sent with every request, such as your conventions, preferred libraries, and plot style. You get up to 2,000 characters, set once instead of repeating yourself.