How billing works: prepaid credit, settled per token

Top up from $10, by card or crypto

Top-ups happen in the console, from $10 to $10,000 at a time. The top-up window takes a card through our payment processor’s form, or crypto through a separate crypto checkout in the same window. Either way the credit lands on the same balance, and your first payment also creates your first key.

Larger top-ups add bonus credit: +5% from $200, +10% from $500, +15% from $1,000. The bonus is credit, not money, so a refunded or disputed payment takes its bonus back with it. Crypto top-ups earn the same bonus as card top-ups.

Each top-up also holds the per-token prices you were paying for 30 days. Inside that window a model’s price can drop for you but not rise; a rise reaches you at most 30 days later. A model added after your top-up bills at its current price.

How a request is priced

Every chat model has an input and an output price per million tokens, and some have a lower price for cached input. A request costs its tokens times those prices, counted to the token, with no per-request fee on top. The price index has every row; these five are read from the same catalog:

A roleplay turn of 1,500 prompt tokens and 500 reply tokens costs $0.0008 on Roleplay Uncensored and $0.0042 on Hermes 4 405B, at the 2026-10-27 prices. The $10 minimum covers about 13,333 such turns on the first and 2,380 on the second.

Prompt tokens a host counts beyond what your request could hold are left out of the charge and out of your token counts. A non-streamed answer carries its settled charge in the x-uplink-charge header; a streamed one ends with its usage block. The console’s usage log lists every request with its id, tokens and charge, and the billing page has a CSV of every request for each month.

When the prepaid LLM API balance hits zero

A request goes ahead only when the balance covers its estimated cost: the prompt, plus the most output it may return. When it cannot, the API answers 402 at once, sends nothing to a model and charges nothing. The body is the OpenAI error shape:

Retrying does not help; a top-up does, and the link in the message opens the top-up form. Without max_tokens, the estimate assumes 16,384 output tokens. On Hermes 4 405B that is $0.0786 for the output alone; with max_tokens set to 500 it is $0.0024. Setting max_tokens lets a small balance still cover a short answer.

Two other 402 codes come from limits rather than the balance:

• monthly_limit_reached: the key hit the monthly spend limit set on it in the console. Raise the limit or wait for the next month.
• spend_cap_reached: a new account may spend up to $10 before its first payment settles.

You hear before you hit zero. An account that made requests in the last 7 days and has no auto top-up gets one email when its balance drops below a day of its recent spend (or $1, whichever is higher), and one more when it runs out. Each top-up re-arms both.

Requests that cost nothing

• Failed requests. A request the host does not answer, or answers with an error, is not charged.
• Refused answers. When the model itself refuses, by its own refusal signal (a refusal field, or a finish reason of refusal or content_filter), the answer is free. The gateway never guesses a refusal from the wording.
• Answers that miss the format you asked for. A non-streamed request that asks for JSON (response_format with json_object or json_schema, per the OpenAI structured outputs guide, or the header x-uplink-require: json) pays only for an answer that parses and matches the schema. A miss is retried once, and the missed answer is free.
• Refused image and video prompts. A prompt the gateway refuses under the acceptable use policy is refused before it is priced. A generation that fails after the host accepted it is refunded.

Limits on the guarantees

Refused and missed answers are free up to $5 per account per UTC day; past that they bill as usual, so neither guarantee can be farmed. An answer cut off by your own max_tokens is billed, because that limit is your choice. A refunded non-streamed answer carries the header x-uplink-refund with the reason.

Auto top-up and the monthly plan

Nothing charges a card except a top-up you make, auto top-up, or a monthly plan you bought. Auto top-up refills the balance from a card saved with our payment processor: when the balance drops under the threshold, the set amount is charged and credited. The threshold runs from $5 to $500 and the amount from $10 to $500. It tries at most once an hour and 5 times a UTC day, and 3 declines in a row switch it off with one email.

Auto top-up can also switch on by itself. If a card is already saved with our payment processor when your first card top-up arrives, it turns on with a $5 threshold and the amount of that top-up (from $10 to $500), and the receipt says so. The console’s billing page saves a card, shows the setting, changes it and switches it off.

Our payment processor’s store also sells a monthly credit plan for people who want a top-up on a schedule. Each renewal credits the balance like any top-up and starts a fresh 30-day price hold. Nothing on this site requires it.

Expiry, refunds and disputes

Credit has no expiry date. Nothing in the ledger removes credit for age or inactivity: the balance goes down for requests you make and for payments that are reversed. A month with no requests costs nothing.

Top-ups are not refundable, as the terms say. A payment that is refunded, or disputed with your card issuer, is reversed in the ledger, and keys may be paused until the balance is positive. The unused balance of an account closed for breaking the terms or the acceptable use policy is not refunded.

What is kept, and for how long

Each request leaves one usage record: the request id, the model, token counts, the charge and how long it took. Next to it sit SHA-256 fingerprints of the request and the response, never the text. Prompts and completions pass through and are not written anywhere, unless you turn on replay for a key, which keeps that key’s request bodies for 7 days.

From checkout we keep the email and the amount paid, as our payment processor or the crypto checkout reports them; card numbers stay with the processor. Usage records and ledger entries are kept as accounting records, also after an account closes. The hosts that run the models receive the request and the response for the calls they serve. The privacy page is the full list.

Questions