ExplainerUpdated November 18, 2026
Zero data retention vs "not used for training"
Zero data retention means the provider keeps no prompt after answering, while "not used for training" only means your prompts won't train its models.
A provider can keep your prompts for weeks without training on them.
The two promises side by side
| "Not used for training" | Zero data retention | |
|---|---|---|
| Covers | Using prompts to train or improve models | Storing prompts and outputs after the response |
| Says nothing about | How long prompts are kept, who reads them | Billing records, stateful features |
| Typical at the labs | On by default for paid API tiers | By approval, request or contract |
| Survives a flag? | Usually | Not always, since flagged content can be kept |
Where the gap shows
| API | Training | Kept by default |
|---|---|---|
| xAI API | Not without permission | 30 days |
| OpenAI API | Not unless you opt in | Up to 30 days |
| Claude API | Not by defaultFeedback you send is an exception | Within 30 days |
| Morph | Not on any tier | Up to 30 days |
| Gemini API | Yes on the unpaid tierNo on the paid tier | 55 days |
"No training, any tier: We do not use your Inputs or Outputs to train, fine-tune, or evaluate our models for any reason"
Morph still keeps inputs and outputs up to 30 days.
morphllm.comExceptions to "no training"
- Google's unpaid Gemini API and Mistral's Free mode use prompts to improve products.
- OpenAI trains on API data if you opt in.
- Thumbs ratings are an exception at Anthropic and Mistral.
- Mistral may train on data from Labs and Preview Models.
See do AI APIs train on your data for every provider.
Which promise to ask for
If you worry about your data training a model, "not used for training" covers it, and if you worry about it sitting on a disk, only a retention period does. Silk Compute does not store your prompts or model outputs. See the retention timeline for the others.