Models and pricing

The models page lists every model this deployment serves, what each one charges per million tokens, and where those rates came from. Read it before you open a thread: a thread pins one model for its whole life, and the grant that pays for it meters against that model's rates.

The page renders the same catalog as GET /api/v1/models, so the CLI and the web cannot disagree about what is served or what it costs.

What each column tells you

• Model — the id you name when you open a thread, and the provider lane that serves it. The default is the model a thread pins when you name none.
• Availability — available, degraded, or unavailable. A lane whose credential is not configured is listed as unavailable rather than hidden, and selecting it is refused by name. No call is ever answered by a different model than the one you asked for.
• Input, Cached input, and Output — US dollars per million tokens. Where a lane declares no cached rate, cached reads are charged at the input rate.
• Basis — where the rates came from, and the id of the rate table a call is priced against.
• Context — the context window, and the maximum output tokens one call can produce.

Reading the basis

A price on this page is only as good as its source, so every lane says what its source was.

• declared — the operator entered the provider's published rates. This is the only basis anything may bill from.
• provisional — rates written to make the system run, or rates whose table this deployment can no longer resolve. A working figure, not a bill.
• unpriced — no rates at all. Calls on this lane record no cost.

An unpriced lane is not a free lane. It is the deployment saying it does not know what a call cost, which is a different fact from the call having cost nothing. So an unpriced lane shows the word Unpriced instead of $0.00, a thread that touched one reports no total spend rather than a total that looks complete, and the lanes that made the total unknowable are named beside it.

That distinction survives everywhere the number goes: the models page, the thread page, GET /api/v1/threads/{id}, and your account balance, which publishes whether it is complete.

Ceilings

A grant carries call, token, and cost ceilings. A cost ceiling can only stop spend it can measure, so a grant on an unpriced lane is bounded by its call and token ceilings instead. The thread page shows both what a grant has spent and what it may spend, and an unbounded ceiling reads as ∞ rather than as a blank.

Reading the catalog from the CLI

The same list is available over the API:

openagents api models

Every entry carries availability and pricing_basis. An unpriced model publishes no pricing block, so a client that checks for the key sees no price rather than a zero.