Skip to main content

FAQ

Know before you start.

Billing, top-ups, endpoints and keys, answered plainly. If something is missing here, the docs go deeper.

Billing

9 questions

  1. From your prepaid USDT balance: text models per input and output token, embeddings per token, images per image and video per second, at the prices on the Models & Pricing page. Costs come from the usage the model provider reports, never from estimates. Cached prompt tokens are billed at the model’s cache rates, and function calling is billed as ordinary input and output tokens.

  2. No. There is no subscription and no minimum spend. Every request is paid from your prepaid balance at the price listed for that model.

  3. Requests are rejected with HTTP 402 before they are processed, so your balance never goes negative and you are not charged. Top up at any time to continue.

  4. The hold on your balance is released automatically and nothing is charged. An error from the provider is only charged for usage it has already reported, for example when a stream fails midway.

  5. Before it runs, every request holds its maximum cost. When it finishes, only the actual cost is charged and the rest returns to your available balance. Set max_tokens so the hold doesn’t get too large.

  6. The outcome of the request could not be recorded right away. The funds for that request stay held for a short time and are then resolved automatically: charged from the usage the provider reported, or released. settled means the request was charged; released means nothing was charged.

  7. Yes, for what was used. The provider still bills the whole generation, so we read its final usage and charge that, capped by the max_tokens you requested.

  8. Only for what was already consumed. If the provider errors, cuts the stream or times out after reporting usage, you pay for that reported usage (input, cache and the output so far), and the request is marked Partial. If no usage was reported, nothing is charged.

  9. A request has to get its first response within 90 seconds; use streaming for long generations, because a stream only has to start within that time. A 408 means the request never reached the provider: nothing is charged and it is safe to retry. A 422 timeout_error means it was already sent: we still read the provider’s late reply and charge only the usage it reports (for images and video: what was delivered), marked Partial; a late reply without usage is not charged. That 422 carries x-should-retry: false so SDKs don’t retry it and pay twice; check the Usage page before sending it again.

Top-ups

2 questions

  1. Generate a USDT deposit address on the Wallet page of the dashboard, then send funds to it. An admin reviews each deposit before it is credited to your balance.

  2. Every deposit is reviewed by an admin, so your balance increases once it is approved, not the moment the transfer lands.

Endpoints & models

7 questions

  1. Models, chat completions, messages, responses, embeddings, images and videos. Only endpoints whose usage can be metered accurately are open to customer API keys.

  2. Yes. The official openai and anthropic SDKs work by changing only the base URL and the API key. The openai SDK uses a base URL that ends in /v1; the anthropic SDK uses the host without /v1. See the Base URL page.

  3. Yes. Claude Code, Codex, OpenCode, Pi, Hermes and OpenClaw read the same base URL and API key settings as the SDKs.

  4. Yes. GET /v1/models lists the models your key can use, in the OpenAI format; prices and API formats per model are on the Models & Pricing page. POST /v1/responses is available for models in the OpenAI format. It is stateless: store is always false, so previous_response_id, conversation, background and retrieving a response by ID are refused, and you send the whole conversation in input with every request. See the Responses page.

  5. Yes, on /v1/chat/completions, /v1/messages and /v1/responses, for tools that your own application runs. Tools that would run on the provider’s side, such as web search or code execution, are refused with a 400 that names them. Tool definitions and tool calls are billed as ordinary input and output tokens.

  6. They are SSE comments, sent about every 15 seconds while the model is quiet so that no proxy closes the connection. The official SDKs ignore them. If you parse the stream yourself, skip every line that starts with a colon.

  7. Every request field has an explicit rule per endpoint and per kind of model. A field that cannot be honoured is refused with a 400 that names it and says why, instead of being dropped silently, so you never pay for a request other than the one you sent. The field tables are on the Chat Completions, Messages and Responses pages.

Account & keys

4 questions

  1. Anyone with a personal Gmail account. Choose Continue with Google: there is no form and no password, and the dashboard is ready on first sign-in.

  2. Create one key per app, see when each was last used and revoke it instantly; a revoked key stops working immediately. Secrets are shown only once and only a hash is stored.

  3. Up to 20 active keys per account. A revoked key stops working from the very next request.

  4. The dashboard is opened with Google only, and only personal Gmail accounts (@gmail.com or @googlemail.com) are accepted: Google Workspace (company or school) accounts and other email providers are refused. If your account was created with another email address, contact support to move it to your Gmail address, then sign in with Google. The first time an existing account is connected to Google, all of its API keys are deactivated and its other sessions are signed out for security (balance and history stay); create new API keys afterwards. The address must match exactly: a Gmail address that differs only by dots or a +suffix is not connected automatically, so contact support. A disabled account is refused with an account-disabled message; please contact support. API keys of a disabled account also stop working and get 401 "Invalid API key."

Start with any balance,
however small.

Create a free accountSign in AI API · pay as you go