Private AI API checklist
This private AI API checklist has ten questions to answer in writing before you send sensitive data.
The first two are how long prompts are kept by default and what it takes to keep none.
The ten questions
- How long are prompts and outputs kept by default?
See the retention timeline for each answer.
- Is zero retention the default, a setting, an approval or a contract?
You get the default until approval comes through. See how to get zero data retention.
- Which endpoints does it cover?
OpenAI and Mistral publish endpoint lists, and stateful endpoints are usually out.
- What do stateful features keep?
OpenAI's Responses API and xAI keep responses 30 days unless you send
store: false. - Is API data used for training on your plan?
Free tiers often say yes. See do AI APIs train on your data.
- Who can read retained prompts, and when?
Google's unpaid tier lets human reviewers read what you send.
- What happens to flagged requests?
Anthropic keeps flagged chats up to 2 years. See abuse monitoring retention.
- What metadata is kept, and for how long?
Unfil.AI keeps IP addresses in usage logs for 30 days.
- Do add-on tools change the answer?
Google keeps grounded prompts 30 days, and abliteration.ai's promise excludes web search and fetch.
- Is the policy public and dated?
A dated public page can be checked and quoted.
How ten APIs answer the first two
| API | Kept by default | Keeping nothing takes |
|---|---|---|
| Silk Compute | Zero data retention | A verified emailNo approval step or contract |
| abliteration.ai | Never stored by defaultWeb search and fetch excluded | Nothing, on by default |
| Featherless | Not logged | Nothing, on every plan |
| Unfil.AI | Up to 24 hoursTemporary processing data | No option stated |
| xAI API | 30 daysAbuse and misuse audits | A console settingTeam-wide, not yet on for every team |
| OpenAI API | Up to 30 daysAbuse-monitoring logs | Prior approvalPlus additional requirements |
| Claude API | Within 30 daysUp to 2 years if flagged | A sales requestCovered Models keep 30 days |
| Morph | Up to 30 daysFree and paid tiers | An enterprise addendum |
| Gemini API | 55 daysPrompts, context and output | Settings, or Vertex AIVertex AI for a guarantee |
| Mistral API | No period published | A written requestAt Mistral's discretion |
The private AI API ranking has the other columns.
Then check your own stack
A provider that keeps nothing doesn't help if your own code keeps the prompt.
- Turn off request and response bodies in your application logs, proxies and API gateways.
- Check tracing and observability tools, which often record full prompts by default.
- Check error reporters, which can attach the request that failed.
- Send
store: falsewhere the API has it, and delete uploaded files when you're done. - Strip keys, tokens and passwords from prompts.