Usage-based billing
Published model prices and measured request costs
KeepRouter uses prepaid USD credit with no recurring subscription. The catalog publishes the current customer rate for each model, and successful requests record measured usage and the resulting charge.
Last reviewed 2026-08-15 · Editorial review: KeepRouter Editorial
Price before signup
Every canonical model page shows the price unit that applies to that modality. Text models commonly publish input, output, and optional cached-input rates per token, displayed per one million tokens for readability. Image models may publish a per-image request price. The model catalog is generated from the same production pricing configuration used for billing.
Charge after measurement
A request charge comes from the usage reported for the completed call and the current configured rate. The Usage view records model, route, status, prompt/input tokens, completion/output tokens, cache count or state when available, latency, and charge. Prompt and completion bodies are not written to the request log.
Prepaid does not mean unbounded
Credit is added through one-time top-ups. There is no monthly plan that renews silently, but an automated application can still consume available balance quickly. Use model-scoped keys, per-key spend limits, explicit output caps, and application concurrency limits. Treat a low-balance or insufficient-credit response as a product state, not a generic retry.
Cached input and response cache are different
Provider-reported cached input is a model billing category: repeated prompt tokens may receive a different rate when the selected provider reports them. KeepRouter's optional response cache is a gateway feature for identical non-streaming requests and is off by default. Do not combine the two when forecasting cost.
Build a workload budget
Static “cost per task” claims are fragile because repository size, conversation history, tool loops, and output length change token usage. Run the same representative task with candidate models. Record input, cached input, output, status, and charge. Multiply the measured distribution—not a single best case—by expected request volume and add a retry/failure allowance.
Reconcile money as a ledger
The console exposes balance and credit-ledger entries separately from request usage. A payment acceptance, credited balance, and consumed request are distinct events. For finance or support questions, reconcile the ledger and request IDs rather than inferring spend from a dashboard total alone.
Frequently asked questions
Is there a monthly subscription?
No. KeepRouter uses prepaid, pay-as-you-go USD credit with one-time top-ups.
Where do current prices come from?
The public catalog and model pages are generated from the production customer-pricing configuration used by the gateway.
Are failed requests always billed?
Billing depends on the completed route and recorded usage. Review the request status and charge rather than assuming every error has the same accounting behavior.
How should I forecast agent cost?
Measure representative tasks with the same context, tools, and output caps, then budget from the usage distribution plus retry allowance.