What your app consumed, per API key: sessions, the layer-tokens your users' browsers ran and the ones the server ran, server busy time, bytes served, and an estimate of the server share at the stated GPU rate. Counted from the session reports the SDK posts when a run() call finishes (nothing is counted with telemetry: false). No prompt or output content is ever sent.
The key stays in this tab's sessionStorage and is sent only to the API as a bearer. There are no accounts in this phase: the key is the identity, and a key created here is shown once.
A layer-token is one transformer block applied to one new token; the server's count includes the language-model head as one block-equivalent per token whenever the server ran it. Hours with nothing are omitted.