Zero prompt retention
A zero prompt retention API does not keep the prompt after the request has been served. Refuseless is one: prompts and outputs are never saved and never used to train models, per its privacy policy. The text goes in, the completion comes out, and nothing is left as a record you or the host can read later. This page explains the phrase, because vendors use three similar ones for three different promises, and mixing them up is how a security review goes wrong.
Is zero retention the same as no training?
| Phrase | What it should mean | What it does not mean |
|---|---|---|
| Zero prompt retention | The prompt and, on a strict reading, the completion are not kept after the response. | It does not, by itself, describe billing records, account email, or abuse metadata. |
| No training | Your prompts are not used to update model weights. | The host can still store the text in a log. Training and storage are different. |
| Ephemeral processing | The text lives in memory only as long as the worker needs it, then is dropped. | It says nothing about a gateway, a support tool, or a debug capture in front of that worker. |
You can have all three, or only one. A host can refuse to train on your data and still keep thirty days of prompts for abuse monitoring, which is the documented default on the OpenAI and Anthropic APIs. A host can drop the prompt and still train on something else. When you compare APIs, quote the sentence they wrote. abliteration.ai's data-handling page says prompts and completions are processed in memory and not stored by default, with web search as the exception. Adverserial's homepage says no GPU request logs and no training on customer prompts. Those are their sentences, not ours. The privacy-first API comparison puts six of them side by side.
What does Refuseless claim about retention?
Prompts and the model's outputs are never saved, and they are not used to train models. They exist only in the cache on the GPUs that process them, for a limited time. The API key, the account, login records, and the token counts required to bill the call are a different matter: a usage charge cannot exist with no record that a call happened. Some requests may be served by third-party model providers. Those requests are sent from Refuseless servers without your IP address, email, account id, or API key, and the providers are bound by their own data policies, which Refuseless cannot guarantee in every case. This page does not describe disk layout, regions, or a deletion timer, because the product has not published those.
What should you ask any API about retention?
- Do you store the prompt body after the response is sent?
- Do you store the completion?
- Do you train on either?
- Do you ever forward the request to another provider, and what identifying data goes with it?
- What non-content metadata do you keep in order to bill and to investigate abuse?
- Is there a second product, such as an audit log, that keeps the text on purpose?
If your program needs a signed DPA, a region pin, or a customer-managed key, that is a procurement document, not a blog post. We have not published one here. The operational fact you can use today is the retention line above, plus the request shape on the API page.
