# Models and thinking effort

Canonical: https://synthcrew.io/docs/models-thinking

## Before a request [#before-a-request]

The model list is discovered from the configured account/provider. A model name in documentation is not a promise that every account has access. Availability, compatibility, permissions and available AI balance are different checks.

1. Open the model control in the composer, or use `/model`.
2. Search the available list and select a model.
3. Choose a thinking effort supported by that model in the following menu.
4. Check the composer label before sending.

You can also open thinking effort directly with `/thinking`. More effort can take longer and use more AI balance; it does not certify a design or guarantee completion.

## Missing or stale choices [#missing-or-stale-choices]

Use **Refresh available models** when appropriate. If discovery fails, keep the draft and inspect the connection or account status. Do not paste an API key into the message, repeatedly resend a request, or assume the app switched to a different model.

If a model offers fewer thinking settings than another, select from the actual supported values. Do not infer a missing option means it is enabled invisibly.

## Plan and permissions are separate [#plan-and-permissions-are-separate]

Changing model or thinking effort does not grant tools, switch an active run's permissions or apply candidate changes. Set [Plan/Agent and permissions](/docs/agent) deliberately for the next request.

The customer Client uses company-managed remote service access when enabled. It does not contain the owner's private Codex accounts. Website email support uses a separate bounded service configuration, not the engineering model chooser.

## Verify the outcome [#verify-the-outcome]

Read activity, inspect actual tool results and verify candidate geometry/files before Apply. A polished answer from a stronger model can still be wrong. When reporting a model issue, include the visible model label, requested effort and safe error wording—not credentials or raw provider tokens.