Costs

Model prices used to estimate what Plans and Sessions cost.

The Costs page at /costs manages the model price catalog. Every Plan and Session reports a Model cost catalog figure alongside whatever the provider itself reported, and this catalog is where that figure comes from — one consistent set of prices applied to every run’s token counts, regardless of which AI tool produced them.

Two cost numbers appear on a run for a reason. The provider’s is what will show up on a bill. The catalog’s prices every run the same way, so runs are comparable across tools — and the gap between the two is itself informative.

The price catalog

The default source is LiteLLM’s published model_prices_and_context_window.json, which is the one catalog covering every tool this app drives.

Control	What it does
Model cost URL	Point at your own catalog in the same JSON shape. Leave it empty to use the default
Save	Stores the URL and re-fetches from it
Use default	Clears the field, which is what restores the built-in source
Fetch latest	Refreshes now rather than waiting for the daily cycle

Underneath, the card reports when prices were last fetched, how many models the catalog holds, and whether the source is the Default source or a Custom source. A Refresh due badge means the catalog is missing or more than a day old.

Prices refresh automatically once a day, and a page load that finds the catalog stale triggers a refresh on read. A fetch that fails is reported inline with the underlying error, and the previous prices stay in place rather than being cleared.

Model prices

The table lists every priced model in the catalog. Prices are shown per million tokens, formatted to keep both cheap and expensive models readable — four decimal places under a dollar, two above it.

Column	Contents
Model	The catalog’s model identifier
Provider	The provider the catalog attributes it to
Input / MTok	Price per million input tokens
Output / MTok	Price per million output tokens
Cache read / MTok	Price per million cached-read tokens
Cache write / MTok	Price per million cache-write tokens
Context	Maximum input tokens

Every column sorts. Names open ascending; prices and context windows open descending, since someone sorting by those is usually looking for an extreme. Search filters the table server-side, and rows load 100 at a time as you scroll, with a Showing n of N count underneath.

A dash means the catalog publishes no value for that field — some entries price input and output but say nothing about caching.

Notes

• A catalog entry with neither an input nor an output price is dropped rather than stored as a row that could never price anything.
• Changing the source URL re-prices future estimates only; figures already recorded on past runs are not rewritten.
• The catalog is shared across the whole control plane — it is not per agent, per repository, or per run.

Related pages

Usage

Token spend collected from every enrolled machine.

Session details

Where the catalog cost appears alongside the provider’s.

Plans

Estimated cost per run in the list.

Settings

Other control-plane-wide configuration.