What a run costs

Every model call is priced from the tokens it used, and the cost is carried up to the run, the trial and the experiment. Cost is a first-class signal here: on a sweep of a hundred trials, “which configuration was best” and “which was affordable” are different questions and you want both answers side by side.

Where a rate comes from

Three places, in this order. The first one that has an answer wins.

	Source	When it applies
1	A rate someone wrote for this workspace	A negotiated or enterprise rate that is not the public one
2	The built-in catalogue	Every model the platform ships a published rate for
3	Zero	Nothing knows this model’s price

There used to be a screen for this, and it was a trap

Rates were once only hand-typed, one row per model, in a Settings screen. Nobody knows what gpt-4o-mini costs from memory, so in practice that table was empty — and every run reported a cost of zero, which the interface showed as if it were true.

The screen is gone. In its place Settings shows which models you can actually reach, which is the question people were really opening it to answer. The hand-written rates still exist for the case they were built for — a rate that is yours and not public — but they are no longer the only source.

When cost is zero

A zero is not “free”. It means nothing in the chain knew the rate, and it is the one number on these screens worth being suspicious of.

The fix is to widen the catalogue rather than ask anyone to type rates again — the catalogue is where what-is-known lives, and it is checked by a test that fails if a model can be picked and has no rate. That test exists because this has gone wrong twice: a provider was added, its models became selectable, and every run through them was silently free.

Two things it still cannot know, and both are declared where they apply:

• Vertex regional endpoints carry about 10 % over the global endpoint. The catalogue holds the global rates, which is what the platform uses by default.
• A model the catalogue has never heard of — one discovered over the network from a provider that publishes its own list — has no rate until someone adds one.

Where cost shows up

• On a run, as the cost of the model calls inside it.
• On a trial, so a sweep can be compared by cost as well as by score.
• In the workspace’s own accounting, when it is spending from a shared allowance rather than its own key.

Providers & API keys The credentials that decide which models a workspace can reach at all.