Previously, models not in the models.dev catalogue — custom base_url providers, locally-hosted models, or any model whose pricing the catalogue simply doesn't carry — were priced at $0 with only a log warning. That made session cost tracking and per-turn cost display unreliable for anyone working outside the catalogued set, and there was no way to correct a wrong catalogue price without patching the agent binary.

This adds an optional cost block to model configuration that lets you declare token pricing in USD per million tokens, covering input, output, cache_read, and cache_write. When present, it overrides whatever the models.dev catalogue would have returned. An all-zero block is valid and means "priced, but free" — a meaningful distinction from being unpriced. The prices feed the existing per-turn cost computation, session cost accumulation, the /model picker display, and the after_llm_call hook's cost field. Validation rejects negative values and disallows combining cost with first_available.

The schema (agent-schema.json) is updated, a worked example lives in examples/custom-pricing.yaml, and the feature is documented under the "Custom Token Pricing" section in docs/configuration/models/index.md and referenced in docs/configuration/hooks/index.md. Tests in pkg/config/latest/cost_test.go cover config parsing and validation; pkg/runtime/cost_test.go covers the runtime cost computation path.