Model routing

Foreman does not use one model. It picks per task, from the models available to you, based on what that task needs — favouring cheaper models that clear the bar so your balance goes further.

Why you probably should not pin a model

Pinning one model everywhere is the most common configuration mistake. It costs more, because a summarisation step runs on a model bought for hard reasoning, and it usually does not improve output, because the router was already choosing something that met the requirement.

The exception is a specific, observed quality problem on a specific kind of task. Then pin narrowly, and revisit it.

Steps

1. Open Settings → Routing.
2. Set your preference — the meaningful axis is cost-optimised versus capability-first.
3. Use Test routing to see which model a given task type resolves to.
4. If a provider was just added, refresh the registry so its models become selectable.

You're done when a routing test returns a model you would have chosen.

Gotchas

• Model availability follows your keys. A model you have no key for is not routable, on any preference.
• Cheaper is not worse per task. The router only picks a model that meets the task's bar.