Choose a model and how hard it thinks
Pick the model and the effort a session runs on, set defaults for yourself or a project, and see how the choice affects cost.
A session runs on a model, and each model can think for longer or shorter on a task. You can leave both alone, because Evergreen starts with sensible defaults. When a task is harder or simpler than usual, you can change them for one session, or set your own defaults.
Pick a model for a session
You can choose the model before you start a session or while it’s running.
- Before it starts. On New, open the model menu in the footer of the brief box. It has a Model row, an Effort row, and a Per phase row. A pick applies right away. If you can use another coding agent, such as Codex or Grok Build, the menu also has a Harness row, and the model rows list that agent’s models.
- While it runs. In the session’s composer, use the Model and thinking depth chip. A model change takes effect at once for a run in progress. A change of depth applies from the next run.
The picker lists the models you can run. With nothing typed, it shows a short list: the default, the newest model of each family, and the ones you’ve picked before. You can search to reach any other model. A 1M badge marks a model that runs with the one-million-token context window, and default marks the one Evergreen starts with.
The models come in families, listed from the strongest to the lightest: Fable, Opus, Sonnet, and Haiku. Which of them you see depends on what you have access to. Unless your workspace picks another, the default model is Sonnet.
At the foot of the picker, a line says when Evergreen last read your model access. If a model you expect isn’t there, choose Refresh your access under Missing a model?.
Set how hard it thinks
Effort, also called thinking depth, sets how deeply the model reasons before it answers.
| Level | What it means |
|---|---|
| Low | Minimal thinking for the fastest answers. |
| Medium | Moderate thinking. This is the default. |
| High | Deep reasoning. |
| Extra high | Deeper reasoning where the model supports it, otherwise the same as High. |
| Max | The most thinking the model supports. |
Use a different model for each phase
A session moves through three phases, and each can run on its own model and effort. Choose Per phase in the model menu to set them.
| Phase | When it runs |
|---|---|
| Planning | Until you approve the plan. |
| Executing | The approved plan, up to an open pull request. |
| Review | CI fixes and answers to review comments. |
A phase you leave on Same model uses the session’s model. You might plan on a stronger model and execute on a lighter one, for instance.
Set defaults
You can set defaults at three levels. A more specific level wins over a broader one, and you can still change the model on any session.
- A session can pick its own model, even if the defaults say otherwise.
- A project can set a Default model, with its own Planning model, Executing model, and Review model, in Setup on the project’s page. It overrides each person’s own default.
- You can set your own in Settings › Defaults, under Models. There’s a Model row, a row for each phase, and Remember my last pick, which makes a model you pick on a session your new default. Use the org’s picks clears your choices.
When none of these says anything, each phase uses the workspace’s pick, or the provider’s default if there isn’t one. An Owner or Admin sets the picks in Workspace settings › Models.
Automations have their own choice too. An automation can use the Owner’s default, the Org’s picks, or a model you choose with Pick a model, which can also set the effort and a model for each phase. See Start sessions automatically.
How the choice affects cost
Note
A model’s own price passes through at the provider’s price, so a cheaper model costs less. Evergreen’s own fee is the same whichever model you pick, and agent time doesn’t depend on the model.
The model is part of the bill, but not the whole of it. A lighter model is a good fit for simple, well-defined tasks. A stronger model or a higher effort is worth it for work where a wrong answer costs more than the extra time. Billing is covered on its own pages, so check there for how usage is charged.
What to try
- Leave the defaults on for most sessions.
- Raise Effort for a hard bug or a design that needs care.
- Lower it, or pick a lighter model, for small edits and chores.
- Set a project default when everyone working there would pick the same model.
Last updated