Building your AI product after you raise: a founder's guide / Build the product

Build the product

Choosing a model provider: what OpenAI, Anthropic and Google keep and use

Before customer data goes into a model, read what the provider keeps, for how long, where, and whether it trains on it. As read on 26 October 2026, OpenAI says API data is not used for training unless the customer opts in, and keeps abuse monitoring logs for up to 30 days. Anthropic's terms say it may not train on customer content, and it deletes API inputs and outputs within 30 days with stated exceptions. Google uses prompts sent to its free Gemini API services to improve its products and keeps abuse monitoring data for 55 days. This page sets out each provider's terms side by side.

Published October 26, 2026. Editorial.

Key takeaways

OpenAI says data sent to its API has not been used to train its models since March 1, 2023, unless the customer opts in, and keeps abuse monitoring logs for up to 30 days by default.
Anthropic's commercial terms say the customer owns its outputs and Anthropic may not train models on customer content. Its privacy center says API inputs and outputs are deleted within 30 days, with exceptions.
Google uses prompts and responses from its unpaid Gemini API services to improve its products, and human reviewers may read them. Paid use needs a Google Cloud project, the account that holds the service, with active billing.
Zero data retention must be requested at both OpenAI and Anthropic, and OpenAI says it needs prior approval, so request it before a customer contract depends on it.
Google says price changes take effect 30 days after they are posted. Anthropic's terms say 30 days after posting, or earlier if the customer otherwise receives notice.

A model provider receives every piece of text your product sends it. That can include customer names, contracts, health records or source code. What the provider does with that text becomes part of what you promise your own customers, and a large customer may ask about it in its security review. This page sets out what OpenAI, Anthropic and Google said in their own documents on 26 October 2026. It is part of the guide on building your AI product after you raise.

These are the providers' own statements about their own services. Terms change, and a contract you sign with a provider can differ from its public pages, so a lawyer should read the terms that apply to you.

The three providers side by side

Question	OpenAI	Anthropic	Google (Gemini API)
Trains on API data?	No, as of March 1, 2023, unless you opt in [1]	Terms say it may not train on customer content [2]	Free tier: used to improve products; paid tier: not used [6]
Default retention	Abuse monitoring logs up to 30 days [1]	Inputs and outputs deleted within 30 days, with exceptions [3]	55 days for abuse monitoring [7]
Zero data retention	Needs OpenAI's prior approval [1]	Requested through sales, turned on per organization [4]	Not covered in the pages read
Where data is processed	Regional processing for the US, Europe and the UAE (listed models only in the UAE) [1]	Inference can be kept in the US; storage at rest US only [5]	May be stored or cached in any country with Google facilities [6]
Price change notice	Not covered in the pages read	30 days after posting, or earlier on notice [2]	30 days after posting [6]

OpenAI

OpenAI's data controls page says: "As of March 1, 2023, data sent to the OpenAI API is not used to train or improve OpenAI models (unless you explicitly opt in to share data with us)" [1]. OpenAI's pricing page also offers lower prices on fine-tuned models to customers who enable data sharing, as the page API or your own model explains.

OpenAI does keep some data. "By default, abuse monitoring logs are generated for all API feature usage and retained for up to 30 days, unless longer retention is required by law, or is reasonably necessary to protect our services or any third party from harm" [1]. Abuse monitoring is the provider's check that its service is not used against its rules.

OpenAI offers two stricter options, Zero Data Retention (ZDR) and Modified Abuse Monitoring, and says: "Currently, these controls are subject to prior approval by OpenAI and acceptance of additional requirements" [1]. Even with ZDR, some features keep data: OpenAI says "Zero Data Retention ineligible endpoints or capabilities may retain application state when used, even if you have Zero Data Retention enabled" [1]. An endpoint is one address of the API, such as the one for stored files. If your product stores files or conversations on OpenAI's side, check that list.

OpenAI lists ten regions for data residency, meaning where data is kept: the United States, Europe (the European Economic Area and Switzerland), Australia, Canada, Japan, India, Singapore, South Korea, the United Kingdom and the United Arab Emirates [1]. Its table marks regional processing, where the model itself runs in the region, as available for the United States, Europe and the UAE only, and in the UAE for a short list of named models [1]. Outside the US there are conditions: "To use data residency with any region other than the United States, you must be approved for abuse monitoring controls, and execute a Modified Retention amendment" [1]. And "Data residency does not apply to system data, which may be processed and stored outside the selected region" [1]. OpenAI's examples of system data include account information and billing information [1].

For health data, OpenAI says that customers who have signed its Business Associate and Healthcare Addendum, and whose organization is set up with the required retention settings, can use "BAA-eligible endpoints" to process protected health information [1]. A BAA is a business associate agreement, the contract used under HIPAA, the US health privacy law, when a supplier handles protected health information.

Anthropic

Anthropic's commercial terms, effective June 17, 2025, say: "Anthropic agrees that Customer (a) retains all rights to its Inputs, and (b) owns its Outputs" and "Anthropic may not train models on Customer Content from Services" [2].

On retention, Anthropic's privacy center article dated July 1, 2026 says: "For Anthropic API users, we automatically delete inputs and outputs on our backend within 30 days of receipt or generation, except:" [3]. The exceptions it lists include services that keep data longer by design, such as its Files API, separate agreements such as zero data retention, and enforcement of its usage policy [3].

Anthropic's developer documentation adds the details. ZDR is requested through its sales team and "is enabled per organization" by the account team [4]. Even under ZDR or a health data arrangement, Anthropic may keep data where the law requires it or where its automated safety systems flag it: "if a chat or session is flagged, Anthropic may retain inputs and outputs for up to 2 years" [4]. Four models, Claude Fable 5.1, Mythos 5.1, Fable 5 and Mythos 5, "require 30-day data retention and are not available under ZDR unless expressly authorized by Anthropic" [4]. Anthropic offers a health data arrangement: "With a signed BAA and a HIPAA-enabled organization, you can use supported API features to process PHI" [4]. PHI is protected health information.

If you reach Claude through another cloud, the processor changes. Anthropic says: "On Amazon Bedrock and Claude Models on Google Cloud (Google managed), the cloud provider is the data processor" [4]. A data processor is the company that handles the data on your behalf, so read that cloud provider's terms as well.

On location, Anthropic's data residency page says "Currently, "us" is the only available workspace geo", meaning data at rest is stored in the US, and that a request can be set so that "Inference runs only in US-based infrastructure" [5]. The page what an AI feature costs to run covers the price of that setting.

Google (Gemini API)

Google's terms treat free use and paid use differently. For unpaid services, "including, for example, Google AI Studio and the unpaid quota on Gemini API, Google uses the content you submit to the Services and any generated responses to provide, improve, and develop Google products and services and machine learning technologies" [6]. The same terms add: "human reviewers may read, annotate, and process your API input and output", and "Do not submit sensitive, confidential, or personal information to the Unpaid Services" [6]. Google's pricing page marks the free tier with "Content used to improve our products" [9].

For paid services, "Google doesn't use your prompts (including associated system instructions, cached content, and files such as images, videos, or documents) or responses to improve our products" [6]. The terms define paid use narrowly: "Your access to Gemini API is a "Paid Service" only when accessing the API through a Cloud Project associated with an active billing account" [6]. An API key, the password a program uses to call the service, from a project without active billing is on the free terms, whatever data it sends.

Google also restricts where free use may serve users: "You may use only Paid Services when making API Clients available to users in the European Economic Area, Switzerland, or the United Kingdom" [6]. And Google lists an invite-only Contributor tier, "offered as a private preview on an invite-only basis", whose table row for use to improve Google's products reads yes [9].

On retention, Google's abuse monitoring page says "Google retains the following data for fifty-five (55) days for the purposes of detecting and preventing violations of the Prohibited Use Policy" [7]. Separately, logs that the developer chooses to keep in Google's tools are "retained for a default maximum period of 55 days", and can be set to expire after 7, 14, 28 or 55 days [8]. On location, paid-service data "may be stored transiently or cached in any country in which Google or its agents maintain facilities" [6].

How much notice a price change gets

The contract also covers price. Anthropic's commercial terms say it "may update the published rates, to be effective the earlier of 30 days after the updates are posted by Anthropic or Customer otherwise receives Notice" [2]. Google's terms say it "may make changes to this pricing from time to time, effective 30 days after they are posted unless otherwise specified" [6]. Under Anthropic's wording a change can take effect sooner than 30 days if notice reaches the customer first. Plan for a price change with a month's warning or less, and keep the product able to move to another model.

What to do before customer data goes in

Write down which provider, which model and which tier the product uses, and the date you read the terms.
On Google, move to a Cloud project with active billing before any real customer data is sent.
If a customer needs zero data retention, request it from OpenAI or Anthropic early, because both require approval.
If you handle health information in the US, sign the provider's BAA before the first patient record is sent.
If customers are in Europe or the UK, check which regions each provider offers and what conditions apply.
Keep this record with your security documents, because the first large customer may ask for it. See security questions from your first large customer.

Reveneau's view

Reveneau's view is that the provider's data terms belong in the first design decision, next to the choice of model. A product whose customers are banks, hospitals or European companies may need regional processing, zero data retention or a BAA, and each of those narrows the list of providers and models. Choose with those needs written down, and test the shortlisted models on your own eval cases. For a finance product, the guide on SOC 2 for a financial software engineering process covers the controls customers may expect.

Reveneau is an AI software development consultancy with offices in San Francisco and Vancouver. Reveneau says AI writes all of its code and every change must pass an eval suite written from the specification. It says it takes responsibility for the whole project through production and after release. To plan which provider your product should use, see AI development at Reveneau or contact Reveneau.

Common questions

Does OpenAI keep the prompts sent to its API?

OpenAI keeps abuse monitoring logs for all API use for up to 30 days by default, unless longer retention is required by law or needed to protect its services or others from harm, according to its data controls page as read on 26 October 2026. It says data sent to the API is not used to train its models, as of March 1, 2023, unless the customer opts in.

What is zero data retention?

Zero data retention (ZDR) is an arrangement in which the model provider keeps no copy of a customer's inputs and outputs, with exceptions each provider lists. OpenAI says its ZDR controls need its prior approval, and some of its endpoints may keep application state even with ZDR enabled. Anthropic says ZDR is requested through its sales team and enabled per organization.

Does Anthropic train on API customers' data?

Anthropic's commercial terms, effective June 17, 2025, say Anthropic may not train models on customer content from its services. The same terms say the customer retains all rights to its inputs and owns its outputs. Those are Anthropic's own terms for commercial customers, and a lawyer should check the version that applies to your account.

How long does Anthropic keep API data?

Anthropic says it automatically deletes API inputs and outputs within 30 days of receipt or generation, in a privacy center article dated July 1, 2026. The exceptions include services that keep data longer, separate agreements and usage policy enforcement. Its documentation adds that a flagged chat or session may be kept for up to 2 years, even under zero data retention.

Does Google use Gemini API prompts to improve its products?

Google uses prompts and responses from its unpaid Gemini API services, including AI Studio and the free quota, to improve its products, and human reviewers may read them, according to its terms effective March 23, 2026. On paid services Google says it does not use prompts or responses to improve its products. Paid use requires a Google Cloud project with active billing.

Can I use the free Gemini API for an app with users in Europe?

Google's Gemini API terms, effective March 23, 2026, say you may use only paid services when making API clients available to users in the European Economic Area, Switzerland or the United Kingdom. Paid use means access through a Google Cloud project with an active billing account. Google also tells developers not to send sensitive or personal information to the unpaid services.

Where is data processed when I use OpenAI's API?

OpenAI lists ten data residency regions, and its table marks regional processing, where the model runs inside the region, as available for the United States, Europe and the United Arab Emirates only, as read on 26 October 2026. Regions outside the US need approval for abuse monitoring controls and a signed Modified Retention amendment. System data such as billing may be stored elsewhere.

Can I keep data in the United States on Anthropic's API?

Anthropic's data residency page says the US is currently the only available workspace geo, which is where data is stored at rest, and that a request can be set so inference runs only in US-based infrastructure. Its pricing page says US-only inference on Claude 4.6 and later models costs 1.1 times the normal price for every token category.

Can I process health data with an AI model API?

Health data can be processed with some AI model APIs once the right contract is signed. Anthropic says that with a signed business associate agreement and a HIPAA-enabled organization, supported API features can process protected health information. OpenAI says customers with its Business Associate and Healthcare Addendum can use BAA-eligible endpoints under the required retention settings. Sign before the first record is sent.

How much notice do AI providers give before a price change?

Google's Gemini API terms say pricing changes take effect 30 days after they are posted unless otherwise specified. Anthropic's commercial terms say updated rates take effect 30 days after posting or when the customer otherwise receives notice, whichever is earlier, so an Anthropic change can apply sooner than 30 days. Plan for a month's warning or less.

Who processes data when Claude runs on Amazon Bedrock?

When Claude runs on Amazon Bedrock, or on Google Cloud where Google manages it, the cloud provider is the data processor, according to Anthropic's API and data retention page as read on 26 October 2026. A data processor handles the data on your behalf. The cloud provider's own terms then decide what is kept and where, so read those terms too.

References

More in Build the product

API or your own model: how a startup should build the AI part

A startup can put AI into its product in three ways: call a model that a provider hosts, through its API; fine-tune a hosted model on its own examples; or run an open-weight model on servers it controls. As read on 26 October 2026, OpenAI is winding down its self-serve fine-tuning and has closed it to new users, and Google's Gemini API has no model that supports fine-tuning. Reveneau's view is that a startup should start with a hosted model, clear instructions and an eval suite, and change approach only when the evals show a gap. This page sets out each option with the providers' own terms.

What an AI feature costs to run, per user, with October 2026 prices

An AI feature costs money every time a user uses it, because the model provider charges for each piece of text sent and received. This page builds one invented example, a support assistant used 40 times a month per user, and prices it with the rates OpenAI, Anthropic and Google published on 26 October 2026. On the mid-priced models the example costs $0.44 per user per month, and on the smallest models $0.022. The page then shows how prompt caching and batch requests change the figure, which charges add to it, which prices are scheduled to change, and the spending limits each provider sets.

Who owns the code when a development partner or an AI writes it

A startup has to own the code it sells, and investors and buyers check that it does. Under US law, copyright first belongs to the author, an employee's work within the job belongs to the employer, and a transfer of copyright from a contractor must be in writing and signed. The US Copyright Office said in January 2025 that purely AI-generated material is not protected by copyright, while creative selection, arrangement and changes to AI output by people can be. This page explains each rule with its source, the Thaler court decision, the UK rule, and what to put in writing with a development partner.

Previous: What an AI feature costs to run, per user, with October 2026 prices Next: Who owns the code when a development partner or an AI writes it