Skip to content
Documentation
User Guide
Switching Models

Switching Models

Mid-session model switching lets you pop between a fast/cheap model and a stronger/slower one without restarting Cogo.

Picker

Type /model with no args:

Model picker (↑/↓ select · enter switch · esc cancel)
▸ gemini-3.1-pro-preview  (current)
  gemini-3.5-flash-preview
  gemini-3.0-pro
  ...

↑/↓ navigate, Enter switches, Esc cancels.

Direct switch

Skip the picker by passing the model ID:

/model gemini-3.5-flash-preview

What changes when you switch

  • The agent is rebuilt with the new model bound, but everything else carries over: project memory, MCP servers, skills, permission state.
  • The conversation context resets — the new model gets a fresh system prompt with the same memory, but no prior turns. (Different models have different context windows and prompt formats; mixing histories tends to confuse them.)
  • A system message confirms the switch:
    Switched to gemini-3.5-flash-preview. Conversation context resets for the new model.

Persistence

If your project has an .agents/config.json, the new model ID is persisted to that file under model.name. Subsequent runs will use the new default.

To switch in-session only (don’t persist), edit config.json afterward to revert.

If you’re running without an .agents/ (built-in defaults), the switch is in-session only — nothing is written to disk.

Why switch?

Common patterns:

  • pro for planning, flash for execution — start a session in gemini-3.1-pro-preview to scope a refactor, then /model gemini-3.5-flash-preview for the long tail of mechanical edits.
  • Cost-sensitive headless runs — set model.name to flash in config.json for CI scripts; flip back to pro interactively.
  • Trying a new release — /model gemini-X-preview to A/B against the current default without committing.
Permissions