OpenAI, Anthropic and Google each say paid API data is not used to train their models by default, but retention, regions and eligibility differ. OpenAI keeps abuse-monitoring logs for up to 30 days and offers Zero Data Retention only after approval. Anthropic deletes API inputs and outputs within 30 days unless an agreement says otherwise. Google keeps Gemini API prompts for 55 days for abuse monitoring, uses unpaid-tier content to improve its products and bars apps aimed at under-18s. Read each vendor's current page before you sign, because these terms change.
The short answer
For a startup on the paid API of OpenAI, Anthropic or Google, each vendor says it does not use your prompts and responses to train or improve its models or products by default. The differences that affect an AI MVP are elsewhere: how long the vendor keeps your data, whether you can get zero retention without a sales conversation, where the data is stored and processed, which tiers may touch personal data, and whether your users may be children.
The trap that catches founders most often is the free tier. Google's terms say that unpaid Gemini API content is used to improve Google products and can be read by human reviewers (developers in the EEA, Switzerland and the UK are covered by the paid-service data terms instead), and that only paid services may be used for apps you make available to users in those places. Prototype with test data, and move to a paid account before real customer data arrives. This page is general information, not legal advice.
The three vendors at a glance
Everything below comes from the vendors' own pages, read on 22 October 2026. Page dates are shown where the page gives one.
| Question | OpenAI API | Anthropic API | Gemini API |
|---|---|---|---|
| Used to train models by default? | No. Not used to train or improve models unless you opt in (since 1 March 2023) | No, for commercial products including the API (page dated 18 August 2026) | Paid services: not used to improve Google products. Unpaid services: used to improve them |
| Standard retention | Abuse monitoring logs for up to 30 days; several endpoints keep application state until you delete it | Inputs and outputs deleted within 30 days of receipt or generation, with exceptions (page dated 1 July 2026) | Prompts, context and outputs kept 55 days for abuse monitoring (page updated 9 June 2026); optional project logs default to a 55-day maximum |
| Zero retention route | Zero Data Retention or Modified Abuse Monitoring, by prior approval | Zero data retention agreements, reviewed per organization | A list of features to avoid; guaranteed zero retention through Google's enterprise platform |
| Regions | Data residency by project in ten regions, with conditions | Traffic may be routed to select countries in the US, Europe, Asia and Australia; data is stored in the US (page dated 15 June 2026) | Prompts may be stored transiently or cached in any country where Google or its agents have facilities |
| Health data | BAA-eligible endpoints under a Business Associate and Healthcare Addendum | BAA for HIPAA-ready services, with a configuration and a request to sales | Terms bar use in clinical practice and to provide medical advice |
| Children | Zero data retention first for personal data of under-13s; safeguards for under-18s | Safeguards, AI disclosure and COPPA compliance for products minors use directly | Terms bar applications directed at or likely to be accessed by under-18s |
OpenAI
OpenAI's data controls page opens with the training position: as of 1 March 2023, data sent to the OpenAI API is not used to train or improve OpenAI models unless you explicitly opt in.
Two kinds of stored data. The page separates abuse monitoring logs from application state. Abuse monitoring logs can contain prompts and responses, plus metadata such as classifier outputs. By default they are generated for all API feature usage and kept for up to 30 days, longer if the law requires it or if OpenAI considers it reasonably necessary to protect its services or a third party. Application state is data an endpoint persists to do its job.
Application state is where the surprises are. In the endpoint table, chat completions and responses show "None, see below for exceptions" for application state and are eligible for Zero Data Retention, subject to limits; the notes under the table add that the Responses API keeps response data for at least 30 days by default, or whenever the store parameter is true. Embeddings, image generation, audio transcription and moderations are also eligible. Conversations, ChatKit threads, agents, the Assistants API, threads, vector stores, files, fine-tuning jobs, evals and batches keep application state until you delete it and are marked as not eligible for Zero Data Retention. If you chose a stateful feature because it saved you code, check what it keeps.
Zero retention needs approval. Zero Data Retention and Modified Abuse Monitoring both exclude customer content from abuse monitoring logs. Both are subject to prior approval and extra requirements, and OpenAI asks you to contact sales. Zero Data Retention also forces the store parameter to false on the two main endpoints. OpenAI reserves the right to make specific models ineligible for either control for specific customers, with written notice.
Regions. Data residency is a project setting, and the page asks you to contact sales to check eligibility. The page lists regional storage for the United States, Europe (EEA and Switzerland), Australia, Canada, Japan, India, Singapore, South Korea, the United Kingdom and the United Arab Emirates, and regional processing only for the United States, Europe and the UAE. Outside the US you must be approved for abuse monitoring controls and sign a Modified Retention amendment, and the UAE needs extra approval. Eligible models released on or after 5 March 2026 carry a 10 percent uplift. If you pick a region without regional processing, OpenAI may process and temporarily store content outside it.
Health data. The page describes a Business Associate and Healthcare Addendum and says BAA-eligible endpoints can process protected health information once your organization is provisioned. It also says web search with live internet access is not HIPAA eligible and not covered by a BAA, while offline or cache-only web search can be covered when used from a Zero Data Retention project.
Children. OpenAI's under-18 API guidance asks organizations serving minors to follow child protection and privacy law including COPPA, and says you should not process the personal data of children under 13, or under the applicable age of digital consent, without first implementing zero data retention. It also lists age-appropriate disclosures, content filters, monitoring and reporting, and age assurance where required, and OpenAI reserves the right to audit.
Anthropic
Anthropic's privacy center separates its commercial products, including the API, from its consumer plans. For the commercial products, an article dated 18 August 2026 says Anthropic will not use inputs or outputs to train its models by default, unless you report feedback or otherwise choose to let it.
Retention. For API users, Anthropic says it automatically deletes inputs and outputs from its back end within 30 days of receipt or generation, except when you use a feature with longer retention under your control (the Files API is the example), when you and Anthropic have agreed otherwise, when it needs to retain content to enforce its Usage Policy, and where the law requires. If automated systems flag content as violating the Usage Policy, Anthropic keeps inputs and outputs for up to two years and classification scores for up to seven. Feedback you submit through a thumbs up or down button is kept for five years. For paid API customers, an article dated 16 March 2026 says Anthropic does not support ad hoc deletion.
A search we ran on 22 October 2026 returned a summary saying Anthropic had cut API retention to seven days in September 2025. We could not confirm that on Anthropic's own pages, and the retention article dated 1 July 2026 says 30 days. Use the vendor page and ask in writing.
Zero data retention. Some API and enterprise coding-tool customers, subject to approval, may have arrangements under which Anthropic does not store inputs or outputs except where needed to comply with law or combat misuse. Anthropic still keeps its user-safety classifier results. Requests are reviewed per organization, so a company with several organizations has to say so.
Covered Models. From 9 June 2026, Anthropic requires prompts and outputs for models it designates as Covered Models to be retained for 30 days to support safety work, on every platform that offers them. Organizations on zero data retention, including through AWS Bedrock, Google Cloud or Microsoft Foundry, have to enable retention before they can use those models. If your security review depends on zero retention, ask which models you are choosing.
Location. The servers article says Anthropic may route customer traffic to select countries in the US, Europe, Asia and Australia, and that data is stored in the US. It points developers to a routing control for the Claude Developer Platform.
Health data. Anthropic offers a business associate agreement for services it calls HIPAA-ready, such as the first-party API with the HIPAA configuration. The page says the organization's primary owner must sign the BAA and then contact Anthropic's sales team to switch it on, that not all API features are covered, and that Covered Models require 30-day retention and are not available with zero data retention unless Anthropic has expressly authorized it for your organization.
Children. Anthropic's guidelines for organizations serving minors say that products letting minors interact directly with its API should add safeguards such as age verification, content moderation and filtering, and monitoring and reporting. Organizations are responsible for complying with child safety and privacy law such as COPPA and should say so publicly, and must disclose to users that they are talking to an AI. Anthropic audits periodically and can suspend an account for a high violation rate left unaddressed.
Gemini API
The Gemini API Additional Terms of Service (last updated 28 April 2026) are the terms most likely to surprise a founder, because the answer depends on whether the project has billing.
Paid or unpaid. Access through Google AI Studio is a paid service when the account can reach a Cloud project with an active billing account or is a Workspace enterprise account, and access to the Gemini API is a paid service only through a Cloud project with active billing. For unpaid services, Google says it uses submitted content and generated responses to provide, improve and develop its products and machine learning technologies, and that human reviewers may read, annotate and process input and output after it is disconnected from your account, API key and Cloud project. The terms tell you not to submit sensitive, confidential or personal information to unpaid services. They add that if you are in the EEA, Switzerland or the UK, the paid-service data terms apply to all services, including unpaid quota.
Paid services. Google says it does not use prompts, system instructions, cached content, files or responses to improve its products, and processes them under its data processing addendum. Its abuse monitoring page (updated 9 June 2026) says Google keeps prompts, contextual information and outputs for 55 days to detect and prevent violations of its Prohibited Use Policy, and that this logged data is used only for policy enforcement. The terms add that the data may be stored transiently or cached in any country where Google or its agents have facilities.
Zero retention. The Gemini Developer API's zero data retention page (updated 14 September 2026) lists what to avoid or switch off: Grounding with Google Search or Maps, which store prompts and outputs for 30 days with no way to disable it; the default state storage of the Interactions API, which you turn off with the store parameter; session resumption on the Live API, which keeps conversation state for up to 24 hours; files, which stay until deleted; and explicit context caching. It says workloads that need guaranteed zero retention or enterprise data processing agreements should use Google's enterprise platform instead. Project logs are opt-in, available only with billing, and default to a 55-day maximum that you can shorten to 7, 14 or 28 days.
Where and who. You may use only paid services when you make an API client available to users in the EEA, Switzerland or the UK. The age rule matters for consumer apps: you must be 18, and you will not use the services as part of a website or application directed towards, or likely to be accessed by, people under 18.
Medical use. The terms say you may not use the services in clinical practice, to provide medical advice, or in any manner that is overseen by or requires clearance or approval from a medical device regulatory agency.
What this means for an MVP
The decisions below are our recommendations, not vendor statements.
- Treat the free tier as test data only. Anything a real customer types should go to a paid project. This is the clearest line in the three sets of terms.
- Match the tier to the data. Write down which data classes your product sends to a model: public content, account details, free text from users, files, health or financial data. Each class should have a named vendor, tier and retention setting.
- Prefer stateless calls if you need zero retention. On OpenAI and Gemini, the stateful features are the ones outside zero retention, so a design that keeps conversation state in your own database gives you the deletion control.
- Check regions before you promise residency. Regional storage and regional processing are different things in OpenAI's table, and Anthropic says data is stored in the US. A line such as "your data stays in the EU" needs a vendor page behind it.
- Decide about children early. The Gemini API terms rule out applications likely to be used by under-18s, OpenAI ties personal data of under-13s to zero retention, and Anthropic expects safeguards. If your audience includes teenagers, that is a model-choice constraint, and the app store rules for children add more (see App Store and Google Play rules for AI apps).
- Sign the agreement before the data arrives. For health data, get the BAA and the configuration in place first, and read the list of covered features. See our guides to building AI with patient data and HIPAA-compliant app development.
- Say it in your own documents. Your privacy policy, your data processing agreements and your app store disclosures should name the model vendors you send personal data to. Apple's guideline 5.1.2(i) asks for explicit permission before personal data goes to third-party AI, and the EU AI Act adds duties of its own (see EU AI Act for AI MVP founders).
- Re-read the pages when you change models or tiers. Retention and eligibility are set per model and per endpoint, and a vendor can change either. If you are still choosing a model, our guide to choosing the right LLM for your MVP covers capability and cost.
Questions to put to a vendor in writing
- Is customer content used to train or improve any model, on this tier and this endpoint?
- How long are prompts and outputs kept, and does that period change with the model?
- Which endpoints and features keep data after the request ends, and how do I delete it?
- Can I get zero data retention for this model, and what does it cost in features?
- In which regions is data stored and processed, and what does "region" cover?
- Which subprocessors can see content, and how will I hear when they change?
- Do you sign a data processing addendum and, for health data, a business associate agreement?
- Which rules apply to apps used by people under 18?
Limits of this page
We read the pages named below on 22 October 2026. They are public summaries and terms, not a contract. An enterprise agreement, a data processing addendum or a cloud reseller such as Azure, AWS or Google Cloud can change retention, regions and eligibility, and a vendor can edit its page at any time. We did not test any vendor's behavior; everything here is what the vendor says. We did not cover other model vendors, open-weight models you host yourself, or the consumer chat products. The summaries and recommendations are ours and are not legal advice.
Sources
All pages were read on 22 October 2026.
- Data controls in the OpenAI platform, OpenAI API documentation
- Under-18 API guidance, OpenAI API documentation
- Is my data used for model training?, Anthropic Privacy Center, dated 18 August 2026
- How long do you store my organization's data?, Anthropic Privacy Center, dated 1 July 2026
- Can you delete data that I sent via API?, Anthropic Privacy Center, dated 16 March 2026
- Data retention practices for Covered Models, Anthropic Privacy Center
- I have a zero data retention agreement with Anthropic. What products does it apply to?, Anthropic Privacy Center
- Where are your servers located?, Anthropic Privacy Center, dated 15 June 2026
- Business Associate Agreements (BAA) for Commercial Customers, Anthropic Privacy Center
- Guidelines for Organizations Serving Minors, Anthropic Help Center
- Gemini API Additional Terms of Service, Google AI for Developers, last updated 28 April 2026
- Abuse monitoring, Gemini API documentation, last updated 9 June 2026
- Zero data retention in the Gemini Developer API, Gemini API documentation, last updated 14 September 2026
- Data logging and sharing, Gemini API documentation, last updated 14 October 2026