Platform

Pricing and billing

Prepaid credit, pay as you go. Each request costs what the model’s price says, and the request log shows exactly what it counted.

Credit

  • Add credit by card under Billing, from $10. It’s shared by every key in the workspace.
  • Credit you buy doesn’t expire while your account is open. Bonus credit from a promotion expires 90 days after it’s added.
  • Auto-recharge, under Billing → Payment method: save a card (it goes to Stripe, never to us) and we add a set amount whenever the balance available for new requests (the balance less what running requests have set aside, see below) falls below your threshold, at most once every two minutes and ten times a day, with a receipt by email. Many requests running at once can start a top-up while the balance itself is still above the threshold. A declined card pauses it, and the owners and admins are told why.
  • Before a request starts, what it may cost (its input, and all the output its max_tokens allows, or a default amount without one) is set aside from the balance until it’s done, so requests running at the same time can’t spend the same credit twice. When the balance, less what your running requests have set aside, can’t cover a request, it’s refused with 402 insufficient_quota; a lower max_tokens asks for less. A streamed request can start with room for its input and about 1,000 tokens of output.
  • A model that may reason where you can’t see it (one that can reason, or any model you ask to) only shows that reasoning in the usage at the end. So when the balance covers the start of such an answer but not all the output the request allows (its max_tokens, or about 32,000 tokens without one), the max_tokens sent to the model is lowered to what the balance covers, with any thinking budget lowered to leave room for the answer, and all of that is set aside. The model stops there, so reasoning it doesn’t show you can’t take the balance below zero. An answer whose output you see as it comes isn’t lowered: it sets aside more as it goes.
  • A long answer sets aside more as it goes. If the balance can’t cover more, or the answer reaches the max_tokens it was lowered to, it stops there: a stream ends with an insufficient_quota error event, and a non-streaming request returns what was written so far with finish_reason: "length". Either way you pay only for what was generated. Rarely, the balance can still end slightly below zero; the next top-up covers that.

What a request costs

KindBilled by
ChatInput tokens and output tokens, at the model’s price per million. Cached input at the cached price where the model has one. On Claude, input written to the prompt cache costs 1.25 times the input price, or 2 times for the one-hour cache, as Anthropic bills it.
Images, PDFs, audio, video in a messageThe input tokens the model counts for them (see Images, PDFs, audio and video)
ReasoningThinking tokens count as output tokens, whether or not the model shows them
Images madePer image, by resolution
VideosPer second, by resolution
EmbeddingsInput tokens only, counted as OpenAI counts them, with no minimum per request. The vectors are free.

Prices are on the models page. Each model’s own page shows the lab’s list price beside ours.

Output tokens are what the model wrote for the answer you got: its text, its reasoning and its tool calls. Never a retry, another attempt, or anything a route ran on its own. If a route reports more output than the model could have written in the time the answer took, we correct the count before billing it, and the usage in the response shows the corrected number.

What’s free

  • Failed requests, and requests refused by a key’s rules or the balance.
  • Images a model declined, and failed videos.
  • Listing models and counting tokens.
  • The instructions we add to keep models answering as themselves: you’re only billed for what you send and receive.

A request that stops partway is billed for its input and the output generated until it stopped: one you cancel or disconnect from, a stream that broke off, and an answer the balance or a key’s budget ran out on. When you disconnect, Zurelay keeps reading the answer for up to 30 seconds, and what the model writes in that time is billed too. On a route that stops when we disconnect from it, the answer stops there instead and you pay only for what was generated before, unless the model may be thinking out of sight (it can reason, and none of its reasoning has come through yet): that thinking is billed up to the moment we disconnect though it never shows, so such an answer is read on as well. With smart routing, a request answered by an alternative is billed at the alternative’s price, and a Zurelay Auto request at the price of the model that answered it.

Rarely, a model sends back an error message in place of an answer. If every attempt at a request ends that way, it’s free like any failed request: you get 503 model_unavailable with retry-after. Only when the model kept answering other customers normally at the same time, so something in the request causes the message (a prompt can be written to), is the input billed once, since the model read it each time. That 503 says so, has no retry-after, and sends x-should-retry: false so the OpenAI and Anthropic SDKs don’t resend it. A message about a balance or quota is never billed.

Where to see it

  • The request log: every request with its model, tokens, cost and speed.
  • Each API key’s spend today and this month on the API keys page.
  • Your payment history under Billing. Stripe emails you a receipt for every card payment.
Questions about a charge? Open a request under Support in the dashboard with the request’s x-request-id.

Questions, or something missing? Ask our support team and we’ll answer by email, usually within a minute.