IllustrationAs of October 31, 2026, OpenAI keeps API abuse logs up to 30 days and does not train on API data by default. Anthropic deletes inputs and outputs within 30 days unless an exception applies. Google logs paid prompts for an unspecified period and uses unpaid content to improve products. Zero retention needs approval or agreed terms.
If your device sends voice or text to a cloud model, your customers will ask one question: is what I say stored, and is it used to train a model? The honest answer lives in three or four vendor pages that use different words for similar things. This page puts the published terms side by side so a product manager or privacy owner can see the differences, the traps, and what to measure on their own system. It is not legal advice, and every row carries the date we read it.
Terms change often. Every figure below was read from the linked official page on October 31, 2026. Re-read the source before you put any of it in a customer answer, a privacy policy or a contract.
How do the three providers compare on retention and training?
Each cell below is taken from the provider’s own page, linked in the row. “Not stated on this page” means the cited page does not say, which is different from “no”.
| Provider | Default retention | Used for training | Zero retention | Terms that extend retention |
|---|---|---|---|---|
| OpenAI API Source: your-data guide, as of October 31, 2026 | Abuse-monitoring logs are generated for all API feature usage and retained for up to 30 days by default. | Data sent to the API is not used to train or improve OpenAI models. | ZDR removes customer content from abuse logs. It requires OpenAI approval. | Audio outputs application state is stored for 1 hour to enable multi-turn conversations. |
| Gemini API Source: API additional terms, as of October 31, 2026 | Paid services: Google logs prompts and responses for a limited period, solely for detecting and preventing violations. The page gives no number of days. | Paid services: prompts are not used to improve products. Unpaid services: Google uses submitted content and responses to provide, improve and develop products, and human reviewers may read it. | Not stated on this page. | Grounding with Google Search stores prompts, context you provide and output for thirty (30) days. |
| Anthropic API Source: data retention article, as of October 31, 2026 | Inputs and outputs are automatically deleted on the backend within 30 days of receipt or generation, with exceptions. | Not stated on this page. A separate Anthropic article covers it; read that one before you answer. | Listed as an exception for ZDR or other agreed terms. Which products it covers is not stated on this page. | Exceptions listed: the Files API, agreed terms such as ZDR, Usage Policy enforcement and legal compliance. |
Two notes on reading the table. The OpenAI endpoint table lists the transcription and translation endpoints as ZDR eligible with no abuse-monitoring or application-state retention. That is our reading of the table, not a sentence on the page, so confirm it against the current page. For Gemini, the same terms page says the paid-services terms apply to users in the EEA, Switzerland and the UK, even for free access.
What do those differences mean for a voice device?
This section is engineering reasoning from the table, not a report of any company’s incident.
- “Not trained on” and “not stored” are separate questions. OpenAI’s default says no training, and also says logs are kept up to 30 days. A customer who asks “is it stored?” needs the second answer, not the first.
- The 30-day window is a ceiling for the logs, not for everything. Features that keep state, such as audio output state for multi-turn use, Grounding with Google Search, and the Files API, carry their own retention.
- Free and paid keys can have different terms. A prototype on an unpaid Gemini key is covered by different terms than a paid production account.
- An unspecified period cannot be repeated to a customer as a number. Google’s paid-tier logging has no stated day count. Say “not specified by the provider”.
- ZDR is a contract step with lead time. At OpenAI it needs approval, so it belongs in the schedule before launch rather than the week of launch.
Which data hops does each term cover?
The vendor terms cover one hop of a longer path. For a wearable that records speech, the path usually has these hops, and your customer’s question applies to all of them.
| Hop | Whose terms apply | What you need to know |
|---|---|---|
| Device storage and buffer | Yours (firmware design) | What is kept on the device, for how long, and whether it is encrypted. |
| Phone app and radio link | Yours, plus the phone platform | Whether audio is cached on the phone and what OS backups may capture. |
| Your backend and logs | Yours (your cloud account) | Request logs, queues and error traces often hold the same content the model saw. |
| Model provider API | The provider terms in the table above | Retention, training use, ZDR and feature-specific storage. |
| Any second provider | Each provider’s own terms | Speech-to-text, translation or search services in the chain each have their own page. |
Our article on keeping a voice wearable off China-hosted cloud covers where that path can be hosted. On our projects the default is that the client holds the keys and the backend runs in the client’s own cloud account, so the model-provider hop is a contract the client signs directly. See how we handle data sovereignty.
What should you measure on your own system?
The vendor pages tell you what the provider says. They do not tell you what your product actually sends. These checks are for your team to run. We have not run them and report no results here.
| What to test | How | How to set the pass line |
|---|---|---|
| Which endpoint and tier each request uses | Log the model endpoint, the project or key ID and the billing tier of every call from a staging build. Include the prototype keys. | Every production request maps to an account whose terms you have read. Any request on an unpaid or personal key fails. |
| What the payload actually contains | Capture outbound requests from your backend in staging with a marker phrase spoken or typed on the device. Search the payload for the marker and for device or user identifiers. | Decide in advance which fields may leave your cloud account. Any extra identifier or raw audio you did not intend to send fails. |
| Which retention-extending features are switched on | List every call that uses search grounding, file upload, stored conversation state or audio output state. | Each feature maps to a row in your retention table with its own period. A feature with no documented period fails until you ask the provider. |
| Your own copies of the data | Send the marker, then search your logs, queues, object storage and error traces for it. | You define your own retention period. The marker must be gone from every store after that period, and absent from stores you did not plan. |
| Your deletion path | Run your delete-my-data flow on a test account and repeat the marker search. | No copy remains in the stores you control. Provider-side deletion follows the provider’s stated window; you cannot verify it from outside, so do not claim you did. |
What should you ask a provider or your counsel before you commit?
- Which exact endpoints will we call, and which row of the provider’s retention table does each one fall under?
- Is our account on a paid tier, and does any key in our system sit on a free tier?
- Is zero data retention available to us, who approves it, and what does it exclude?
- Do any features we use store data longer than the default window?
- For Gemini paid usage, what is the logging period, since the terms give no number?
- For Anthropic, what does the training-policy article say, and which products does ZDR cover?
- Where do we promise customers a retention period that is shorter than a provider’s stated window?
- Do voice features disclose recording to the person speaking, and is consent captured for adults and elders? Never frame a product as covert recording.
- Who rechecks these terms, and on what schedule?
How do you answer “is my voice stored or trained on?”
A copy-ready structure that stays true to the sources:
- Say which provider and which tier handles the request.
- Give the provider’s stated default retention, with the date you read it.
- Give the provider’s stated training position, or say you have not confirmed it.
- State which retention-extending features your product uses.
- State your own retention and deletion behavior for the other hops.
- Add a line that the terms were read on a specific date and are rechecked on a schedule.
Questions about LLM data retention
Does an LLM provider store the audio my device sends?
It depends on the provider, the endpoint and your account tier. OpenAI lists abuse-monitoring logs kept for up to 30 days by default, and lists its transcription and translation endpoints as eligible for zero data retention. Anthropic deletes inputs and outputs within 30 days by default. Google logs paid-tier prompts for a limited period it does not quantify. Read the table above and the provider pages, because terms change.
Is my customers’ data used to train the model?
OpenAI states that data sent to its API is not used to train or improve its models. Google states that paid services do not use prompts to improve products, while unpaid services do, and human reviewers may read that content. The Anthropic page we cite covers retention only, so we state no training position for it. Check Anthropic’s separate training article before you answer a customer.
What is zero data retention, and can I just turn it on?
Zero data retention (ZDR) removes customer content from the provider’s abuse logs. At OpenAI it requires approval from OpenAI, so plan for a request and a wait. Anthropic lists ZDR as an agreed-terms exception to its 30-day deletion. Ask each provider what your account is actually covered by, in writing.
Does the free tier of a model API change the answer?
For Google it does. Its terms say unpaid services use submitted content and responses to provide, improve and develop products, and that human reviewers may read it. Paid services do not use prompts to improve products. A prototype built on a free key can therefore carry different data terms than the production account.
Is this page legal advice?
No. It compares published vendor terms as of October 31, 2026. Whether a given setup meets GDPR, CCPA, BIPA or any other law is a question for your counsel, using the regulator’s own text. Tell us about your device and we will map the data path with you.
Primary sources
- OpenAI, Data controls in the OpenAI platform: developers.openai.com/api/docs/guides/your-data. Accessed 2026-10-31.
- Google, Gemini API Additional Terms of Service: ai.google.dev/gemini-api/terms. Accessed 2026-10-31.
- Anthropic, How long do you store my organization’s data: privacy.claude.com. Accessed 2026-10-31.
Building a device that sends speech to a cloud model? Tell us about your device and we will map the data path hop by hop. For the product side, see AI recording wearable development.
Planning an AI recording wearable? Tell us what it needs to do. A Duobay engineer replies within 1 business day. Send a project brief →