What is kept, and what isn’t
| Thing | Kept? | For how long |
|---|---|---|
| Your prompt, system prompt and history | Never | — |
| The model’s reply and its thinking | Never | — |
| Pictures you send | Never | — |
| One row per call: time, key, model, tokens, credits, status | Yes | 30 days |
| Your key | Only its SHA-256 fingerprint and last four characters | Until you delete the key |
| Your account | An account key, with no email, name or password | While the account exists |
| A fingerprint of each message of a conversation’s last call (GLM only) | In memory, to bill repeated starts as cached | 10 minutes |
| A card payment | Kept by our payment processor, as a payment needs | The processor’s own terms |
The console shows your call rows exactly as they’re stored and exports them as a CSV, so you can check the list above against your own account. Field by field: What we keep.
Where the model runs
A hosted model has to read your prompt to answer it, so privacy depends on the host as much as on us. We only serve a model when its host keeps nothing:
- DeepSeek V4.1 Flash Uncensored runs on a private tier that stores no prompts or replies.
- GLM 5.3 Flash Uncensored runs on a host that keeps zero data: nothing is stored once the reply is sent. Its GPUs belong to independent operators, which is worth knowing if your threat model includes the machine itself.
Neither host is told who you are. Calls reach them from us, under our account, and our cache key mixes in a hash of your account so one customer’s calls can never share another’s cache.
How that compares with typical API policies
Read any hosted AI API’s data policy and you’ll usually find some of these. They’re common, not universal, and each provider words them differently:
| Common practice | Here |
|---|---|
| Prompts and replies kept for a set period, often around 30 days, for abuse review | Not kept at all. The hard-limit check reads the request and stores nothing. |
| Requests used to improve models, unless you opt out, especially on free tiers | Never used for training. There is nothing stored to train on. |
| Zero retention offered only on an enterprise plan, after an application | The default for every key, from a $5 top-up. |
| Request bodies written to logs for debugging | Bodies are never logged. Errors are recorded as codes and counts. |
| An account tied to an email, a name and often a phone number | An account key and nothing else. |
Zero retention is not encryption from the host: the model still reads your words while it answers. If no machine but yours may ever see the text, run a model locally. What zero data retention means covers the difference.
What you can check yourself
- Send a request with no system prompt and ask the model to repeat everything above your message. It finds nothing: we add no hidden instructions.
- Open the console’s call log. Each row has counts only; there is no field for the words.
- Read the
x-nokeep-ignoredheader on a reply: any field we didn’t pass on is named there, so nothing is changed quietly.
Questions
Do you log prompts for abuse monitoring?
No. An automated check reads each request against the hard limits before it runs and returns a yes or no. What it read is not stored, and no person reviews requests.
Is my data used to train models?
No. We store no prompts or replies, and neither host does, so there is nothing to train on.
Is the API end-to-end encrypted?
Calls travel over HTTPS, but the model has to read your prompt to answer it. Zero retention means nothing is kept after the reply, not that nothing reads it. For text no other machine may see, run a model locally.
Can I delete the call rows sooner?
They delete themselves after 30 days. They hold counts, not content: the time, key, model, tokens, credits and status.
Do I need to give an email to get a key?
No. Buying credits makes an account key; there is no email, name or password. A card payment is handled by our payment processor, which keeps what a payment needs.
Is it also uncensored?
Yes. Both models have their refusals removed from the weights, with four hard limits kept. See the uncensored LLM API page.