Zero data retention

We don’t store your prompts or outputs.

That holds for every request you send with a Tokens.Pay key: content exists in memory while the request runs, then it is gone. Below is exactly what that covers, the exceptions, and the billing record we do keep.

Prompts are not stored: AI Runtime, the gateway your key calls, does not write prompts, messages, files, images, audio or embedding inputs to disk, a database, logs or analytics. They exist in memory only while the request runs.
Outputs are not stored: Completions, generated images, speech and transcripts go back to you and are not kept.
No training, no human review: Your content is not used to train or evaluate models, and no person at Gridheap reads it.
Providers see us, not you: AI Runtime rebuilds each request field by field before it goes to the model’s provider, with no name, email, account or key of yours in it. To the provider, the customer is AI Runtime.
We keep the bill: Each request leaves a billing record: id, time, model, token or unit counts, cost, status and the provider used. No IP address or user agent is stored with it.

The path a request takes

1 · Your code

Your app or SDK

Sends the request with your rk_ key to api.airuntime.dev.
2 · AI Runtime

Memory only.

AWS us-east-1. Checks the key, reserves the cost, forwards the request. Keeps the billing record, not the content.
3 · Model provider

Runs the model

Under its own API data terms. Sees AI Runtime as the customer, not you.

Tokens.Pay sells the credit and shows your usage. Your API calls never pass through it, so it never sees a prompt or an output.

The exceptions, exactly

“Zero data retention” means we don’t store your content. It is not a promise about anyone else.

The model provider. Each provider processes your request under its own API data terms. Some keep requests for a limited time to monitor abuse. We have no zero-retention agreement with any provider, so we don’t promise one; each provider’s policy is linked below.
Prompt caching. When a provider serves part of your prompt from its cache (billed at the lower cached-input rate), that cache is the provider’s, under its terms. AI Runtime keeps no prompt cache.
Two AI Runtime features Tokens.Pay does not sell. Async GPU jobs and batches (Inference.Pay) store input and output encrypted with AES-256-GCM for a retention you choose, 60–900 seconds, so the result can be fetched; the input is erased once a GPU picks the job up. OCR holds results in memory for at most 20 minutes, keyed by a hash of the document, never on disk. Neither endpoint is on a Tokens.Pay key.

What we keep: the bill

One record per request, so your balance and usage are exact.

The billing record AI Runtime keeps for each request
Field	Why
Request id and timestamps	To find a request when you ask about it
Model and route	To bill the right rate
Token, second or unit counts	To compute the cost
Cost	Your balance and usage history
Status or error code	To tell a success from a failure
The provider used	To reconcile with the provider’s bill
A one-way SHA-256 fingerprint of the request	For idempotency and disputes; it cannot be turned back into the prompt

Billing rows are append-only and kept as financial records. Application logs carry ids, codes and timings only, never request bodies, and are kept 7 days. Error reports have messages, URLs, headers and bodies stripped. Database statement logging is off.

Who processes your requests

Only the providers whose models Tokens.Pay sells. Each runs the model under its own terms.

Who processes requests for each provider’s models
Processor	For	Where	Data terms
OpenAI	OpenAI models (10)	Not stated	Policy
Fish Audio	Fish Audio models (3)	Not stated	Policy
Moonshot AI	Moonshot AI models (2)	mainland China	Policy
DeepSeek	DeepSeek models (1)	mainland China	Policy
Fireworks AI Fallback	DeepSeek models (1)	United States	Policy
DeepInfra Fallback	DeepSeek models (1)	United States	Policy

The full list of companies that process data for Tokens.Pay, from hosting to payments, is in the subprocessor list.

What this site keeps about you

Your Gridheap Account id and email, to sign you in and to tell you about your payments.
Your top-ups: amount, fee, status and the payment provider’s reference. Card details stay with Stripe or PayPal.
Never your API requests: your code calls api.airuntime.dev directly, so Tokens.Pay never sees a prompt or an output.

Deleting your workspace or account revokes its keys and erases stored credentials. Database backups are kept for 7 days, then the data is gone from them too. To delete your account, write to support from the email you sign in with.

Data questions

Does Tokens.Pay store my prompts or outputs?

No. AI Runtime, the gateway your key calls, holds request content in memory only while the request runs and does not write it to disk, a database, logs or analytics. Tokens.Pay itself never sees your API requests.

Do model providers keep my data?

Each provider processes your request under its own API data terms, and some keep requests for a limited time to monitor abuse. We have no zero-retention agreement with any provider. Their policies are linked on this page.

Is my data used to train models?

Not by us: your content is not used to train or evaluate models, and no person at Gridheap reviews it. What a provider may do is set by its own terms.

Which requests are processed in China?

Requests for Moonshot AI models and DeepSeek models go to the provider’s own API in mainland China (a fallback, where listed, may run elsewhere). The table on this page gives every provider’s region where it is known.

The legal documents

This page explains; the Privacy Policy is the binding text.

Privacy Policy the legal text, in the Gridheap Legal Center
Subprocessors every company that processes data for Tokens.Pay
Terms including refunds and acceptable use

Last reviewed 2026-11-07. Security details are on the security page.

Get one key for every model.

From $5, plus a 5% top-up fee (minimum $0.79). No subscription; paid credit never expires.

Add credit