Usage metering

How tokens, requests, and spend are counted.

Every successful request is metered. Counts and spend roll up to the Usage view on your dashboard and to GET /v1/management/credits.

What’s counted

Field	Source
`request_count`	One per `/v1/chat/completions` call
`prompt_tokens`	The model’s `usage.prompt_tokens`
`completion_tokens`	The model’s `usage.completion_tokens`
`cached_tokens`	Cached-prefix portion of the prompt, when the upstream reports it
`cost`	Per-token price applied to the above; same value returned inline as `usage.cost`
`status_code`	HTTP status of the upstream response

Cost reporting

Every successful response includes usage.cost (USD) inline — the exact amount debited from your balance — plus an X-Dynoyard-Cost-USD response header on non-streaming calls. The dashboard Spend view and GET /v1/management/credits reconcile to the same number to the cent.

Streaming

For SSE streams we automatically set stream_options: { include_usage: true } (when you haven’t) so the final data: chunk before [DONE] carries the usage object including usage.cost. If you explicitly set include_usage: false, token counts may be unavailable for that request — keep the default unless you have a reason to disable it.

What’s NOT counted

Failed requests (4xx / 5xx) — recorded for diagnostics, never billed.
GET /v1/models and other non-completion endpoints — free.

Metering reliability

Usage events are recorded fire-and-forget — no synchronous round-trip on your request path. A periodic reconciliation pass settles any gap as a safety net, so billing stays accurate even if an individual telemetry write is dropped.

Billing

Pricing is pay-as-you-go per-token — see billing and the pricing page. No tiers, no subscription.