Pradra StudioDocs
Sign in

Budgets & costs

Generative AI calls cost real money per token. Pradra Studio meters that spend on every call and lets you set monthly budgets that stop the spending — hard — when the limit is reached. Classic ML predictions carry no per-call provider cost and are not metered.

What is tracked

Every generative prediction records its input tokens, output tokens, and cost in USD, priced from the model's configured rates. You'll see the cost:

  • per call, in the deployment's prediction history (a Cost column);
  • per capability and per organization, aggregated by month, on the Budgets page.

Fine-tuning jobs draw from the same monthly pool — and their cost is booked even if the job fails, because the provider charged you either way.

How budgets behave

A budget caps a calendar month (UTC) in USD, tokens, or both, at two levels:

LevelCaps
OrganizationAll generative spend across the org.
CapabilityOne capability's spend.
  • At 80% of a budget, the spend bar turns amber and an alert fires — your cue to raise the limit or investigate the traffic.
  • At 100%, further generative calls for that scope are refused with a clear "rate limited" response (429) until the month rolls over or the limit is raised. A budget is a hard stop, not a suggestion.
  • Budgets reset implicitly on the first of each month.

The consumer-facing failure is graceful: the calling application gets a standard error envelope it can handle, not a hung request.

Working with budgets in the Studio

  1. Go to Budgets in the sidebar (Operations group).
  2. The organization budget card shows the org-wide limit and this month's spend.
  3. The capability budgets table lists each capability's budget with a spend bar — amber at 80%, red at 100%. Add a budget by picking a generative capability that doesn't have one yet.
  4. A read-only summary of the same numbers appears on each generative capability's page, with a link back here — Budgets is the only place limits are edited, so there is one source of truth.

Choosing sensible limits

  • Start with a generous org budget as the backstop, then add capability budgets only for the ones exposed to unpredictable traffic.
  • Set the limit from expected monthly volume × price per call — the prediction history's cost column tells you the real per-call number after a few days of traffic.
  • Treat an 80% alert mid-month as data, not noise: either the feature is more popular than planned (raise the limit deliberately) or something is calling it that shouldn't be (check the history).

Next