Registering Models

A model is the raw capability behind an agent — the LLM it thinks with. The model registry is where you tell Circus which models exist and what they cost, so agents can be pointed at them and so every run can be priced accurately.

What you register

Each model entry records the exact provider-facing model name, its provider (OpenAI, Anthropic, Google, or xAI), the base API URL, and the cost per one million input tokens and per one million output tokens. The name must be unique, the URL must be valid, and the costs must be zero or positive. Pricing here is base pricing — standard published per-token rates; cached, batch, and other discounted tiers aren't modelled in this version.

Why the pricing matters

These token prices aren't cosmetic — they're what Circus uses to compute the real cost of every agent step. When a workflow reports how many tokens an agent used, the platform multiplies by the model's current price to get the dollar cost that shows up in your logs and analytics. (Cost is always calculated from current registry pricing at the moment a step is logged; changing a price affects future calculations only, and never rewrites history.)

Disabling a model, and the ripple

Each model has an enabled/disabled status, and disabling one is not a local action — it cascades. Any agent assignment that uses the model is disabled, any workflow configuration containing that assignment becomes invalid, and any workflow depending on that configuration becomes unavailable until the dependency is resolved. Re-enabling the model reverses the cascade, provided nothing else in the chain is still disabled. This is the general rule for the whole configuration chain — see The Dependency Chain & Disable Cascades.

The registry also shows you, per model, which agents and workflows are associated with it — so you can see what a change will affect before you make it.