Credits & billing

Credits are what you spend on the platform. Here's how estimates, holds and charges relate.

Estimates are worst-case, and carry no markup

Anywhere you see an estimated cost before you run something - a test, a chat message, a quality check - it is shown in credits, calculated from the platform's own worst-case cost for that run. It is never shown in dollars, and it never includes any markup: it's the same number the real credit hold is sized from.

If your current balance is below an estimate, you'll see a note before you run so a run never stops part way through as a surprise.

Reserve, then settle

Before a run starts, credits are held against your balance for the estimated worst case. Once the run finishes, you're charged what it actually cost - which is usually less than the estimate, since the estimate assumes the longest possible answer from every model. The difference between the hold and the real charge is released back to your balance. If a run can't reach any provider at all, the whole hold is released and nothing is charged.

What a failed model call costs

Each model call in a test is independent, so one provider being down, rate-limited, or rejecting your prompt doesn't stop the others from completing. A failed call is never billed - check the error on that result card for the provider's own reason.

Buying credits

Credits are purchased through Paddle, which also handles tax (including EU VAT) as the merchant of record. Markup is applied once, at the point you buy credits - how many credits a given purchase gets you, and how many models you can run per test, both depend on your plan.

• 10,000 credits equal $1 of real model usage, before any plan markup is applied at purchase.