Model routing

AICoder drives the real claude and codex CLIs, so a session spends your subscription quota. Model routing is the main lever for keeping that cost down.

Defaults

Work	Default model	Why
Coding	Sonnet	Best coding model at a reasonable cost
Reviews, judge, explain, Azure	Haiku	Cheap, ~90% of Sonnet's quality for aux work
Opus / Fable	Opt-in per task	Deepest reasoning, but far more expensive

The CLI's own default model (often Opus) is overridden per job so you don't accidentally run everything on the priciest model.

Overriding per task

In the task drawer you can pick the model for a task:

• Auto — let AICoder route to the cheapest capable model.
• Haiku / Sonnet — fixed choice.
• Opus / Fable — opt-in; shows a warning before it runs.

The expensive-model warning

Choosing Opus or Fable pops a warning ("burns tokens faster…") and, on start, a toast confirming the choice. It's a deliberate speed-bump so cost is always a conscious decision, never an accident.

Configuring defaults

Set the default Claude and Codex models, and the auto-vs-manual routing mode, in Settings → AI models.

Why it matters

Running many parallel sessions multiplies cost fast. Keeping coding on Sonnet and everything auxiliary on Haiku is what makes heavy parallel use affordable.