Configure light models

Use cheaper models for internal tasks to reduce cost while keeping the primary model for code generation.

How cost routing works

Cantrip categorises tasks by how much model capability they need:

Task type	Model used	Examples
Code generation	Primary	Writing charm code, designing integrations
Research	Light	Summarising docs, web search analysis
Test aggregation	Light	Parsing test output, deciding pass/fail
Log queries	Light	Analysing Loki logs, Tempo traces
Context compaction	Light	Summarising long conversations

Light model from the same provider

The simplest setup uses a cheaper model from the same provider:

$ cantrip --provider gemini --light-model gemini-2.0-flash

Or with Claude:

$ cantrip --provider claude --light-model claude-haiku-4-5

Light model from a different provider

For maximum cost savings, use a local inference snap for light tasks while keeping a cloud model for the primary work:

$ cantrip --provider claude \
    --light-provider inference-snap \
    --light-snap nemotron-3-nano

This sends code generation to Claude but runs research, compaction, and log analysis locally with no API cost.

Automatic detection

If you omit --light-model, Cantrip automatically selects a lighter model from the same provider when one is available. You only need to configure it explicitly when you want to override the default choice or use a different provider.