Several of the terms that answer this question changed during 2026, so older summaries are often wrong, and most of them answer only one part of the question. Whether ChatGPT or Claude trains on your company's data is the easiest of the three questions to answer, and the least useful on its own. The terms that decide your exposure are how long your prompts are kept, and who can read them.
This is a dated reference for the people who have to sign off on AI use: CTOs, COOs, and whoever answers to a client or a regulator for where company data goes. Every claim below cites OpenAI's or Anthropic's own pages, as read on 30 September 2026. It is not a case against either provider. Both publish unusually specific commitments, and both deserve to be read accurately.
What Is the Short Answer?
On business plans and APIs, neither OpenAI nor Anthropic trains on your content by default. On consumer plans, the answer depends on a setting, and it is the individual user's setting, not the company's.
Three things are still true on every commercial tier. A copy of your prompt usually exists on the provider's systems for up to 30 days. A narrow set of people at the provider can read it in stated circumstances. And removing that copy entirely, through zero data retention, is an approval process with its own exceptions, not a checkbox.
If you want the broader architecture question, which workloads belong on a rented model and which on your own, we cover it in our guide to running an LLM on-prem versus an enterprise API. This article goes one level deeper, into what the provider terms actually say.
Why Are Training, Retention and Human Review Three Different Promises?
Because they protect against three different things, and a provider can make one without the others.
- Training decides whether your data shapes a future model that other customers use. A no-training commitment protects you from your material resurfacing, in some form, in someone else's output.
- Retention decides whether a copy of your prompt and the model's answer exists on the provider's infrastructure after the request, and for how long. That copy can be breached, subpoenaed or retained under a legal hold, whether or not anyone trains on it.
- Human review decides whether a person at the provider, or a contractor, can read that copy. Every provider needs some route for this, because abuse investigation depends on it.
Marketing pages tend to lead with the first promise because it is the strongest. The second and third are where the conditions live, and they are what a client's security questionnaire will actually ask about.
Not sure which of your teams are pasting company work into consumer AI accounts? Five questions, under a minute, and you will know where to look first.
Find Your AI Ops Score →What Do OpenAI and Anthropic Actually Promise, Plan by Plan?
Last verified: 30 September 2026. Each cell links to the provider page it was taken from. Where a page does not say, the cell says so.
| Plan | Training on your content | Retention of prompts and outputs | Human review |
|---|---|---|---|
| ChatGPT Free, Go, Plus, Pro consumer | May be used to train unless you turn off "Improve the model for everyone". Temporary Chats are not used. (OpenAI) (Data controls) | Saved chats stay until deleted, then permanent deletion within 30 days, with de-identification and legal exceptions. Temporary Chats kept up to 30 days for safety. (OpenAI) (Data controls) | Not specified on the consumer pages we checked. Feedback you submit may be used for training even after opting out. (Data controls) |
| ChatGPT Business, Enterprise, Edu business | Not by default. Only data you explicitly opt in to share, for example through feedback. (Enterprise privacy) | Admins set retention. Deleted conversations removed within 30 days unless law requires otherwise (Business adds: or to protect services from harm). (Enterprise privacy) | Enterprise: authorized employees only for incidents, recovery with your permission, or law. Business: employees for support, abuse and legal compliance, plus contracted abuse reviewers. (Enterprise privacy) |
| OpenAI API developer platform | Not used to train unless you opt in, since 1 March 2023. (Your data) | Abuse-monitoring logs up to 30 days. Stored objects follow their own rules, such as Responses stored for at least 30 days when stored, and Assistants objects kept until you delete them. (Your data) | Access limited to authorized employees for support, abuse investigation and legal compliance, and contracted abuse reviewers. (Enterprise privacy) |
| Claude Free, Pro, Max consumer | Used only if you allow it in your settings, plus conversations flagged for safety review, which may train safety systems. (Anthropic) | Deleted chats removed from back-end storage within 30 days. Up to 5 years, de-identified, if you allow training. Flagged content up to 2 years. (Anthropic) (Aug 2025 update) | Employees cannot access conversations by default, except feedback you share, or designated Trust and Safety staff on a need-to-know basis. (Anthropic) |
| Claude Team, Enterprise Claude for Work | Not permitted: "may not train models on Customer Content from Services". (Commercial Terms) | Kept in the product under your organization's retention settings. Deleted chats removed from back-end storage within 30 days. (Anthropic) | Flagged content is retained, up to 2 years, for Usage Policy enforcement. For Covered Models, human review only through a controlled, logged access path. (Anthropic) (Covered Models) |
| Claude API developer platform | Not permitted under the Commercial Terms, and retained data "never used for model training" without express permission. (API docs) | Inputs and outputs deleted within 30 days, except longer-retention features you control, ZDR agreements, Usage Policy enforcement and law. (Anthropic) | Same as Claude for Work: flagged content retained up to 2 years; for Covered Models, human review only through a controlled, logged access path. (API docs) (Covered Models) |
Read the table by column, not by row. The training column is nearly uniform across business plans and APIs. The retention and human review columns are where the plans, and the providers, actually differ.
Does ChatGPT Enterprise Use Your Data for Training?
No, not by default. OpenAI's enterprise privacy page, updated 8 January 2026, states "We do not train our models on your data by default" for ChatGPT Business, Enterprise, Edu, Healthcare, Teachers and the API Platform, and that data is used for training only if you have "explicitly opted in", for example through feedback mechanisms (OpenAI, Enterprise privacy).
The same page is precise about the other two promises. Enterprise workspace admins control how long data is retained, and deleted conversations are removed within 30 days unless OpenAI is legally required to keep them. Authorized OpenAI employees access Enterprise conversations only to resolve incidents, to recover conversations with your explicit permission, or where the law requires it.
ChatGPT Business, the plan for smaller teams, has the same training default and a slightly wider access clause. OpenAI lists authorized employees for engineering support, abuse investigation and legal compliance, plus "specialized third-party contractors" bound by confidentiality, solely to review for abuse and misuse. If your clients ask whether third parties can see your data, that sentence is the answer for Business.
Two details are easy to miss. OpenAI states it may run business data through automated classifiers and safety tools, producing metadata that does not contain the data itself. And a thumbs-up or thumbs-down is feedback: on consumer ChatGPT, OpenAI says the whole related conversation may then be used for training, even if the user had opted out (OpenAI, Data controls).
Does the Claude API Train on Your Prompts?
No. Anthropic's Commercial Terms, effective 17 June 2025, state that "Anthropic may not train models on Customer Content from Services" (Anthropic Commercial Terms). Its API documentation adds that retained data "is never used for model training without your express permission" (Anthropic, API and data retention).
Those terms cover the Claude API and Claude for Work, meaning the Team and Enterprise plans. When Anthropic changed its consumer terms in August 2025, it stated explicitly that the change did not apply to services under its Commercial Terms, including Claude for Work and API use through Amazon Bedrock or Google Cloud (Anthropic, 28 August 2025).
The consumer side is where the difference sits. Since that change, Free, Pro and Max users choose whether their chats and coding sessions are used to improve Claude. If they allow it, Anthropic may keep that data, de-identified, for up to 5 years; if they do not, the existing 30-day retention period continues (Anthropic, 28 August 2025; Anthropic, consumer retention). Conversations flagged for safety review may also be used to improve Anthropic's detection systems, whatever the setting (Anthropic, model training).
The business terms protect work done inside the business workspace or through the API. They do nothing for an employee who pastes the same material into a personal account, which runs on consumer terms and that employee's own training setting. An enterprise licence is a control only if it is also the most convenient tool in the room.
How Long Do OpenAI and Anthropic Keep API Prompts?
By default, up to 30 days at both. OpenAI generates abuse-monitoring logs for all API usage, which "may contain certain customer content, such as prompts and responses", and retains them for up to 30 days unless the law requires longer (OpenAI, Your data). Anthropic automatically deletes API inputs and outputs within 30 days of receipt or generation (Anthropic, organization retention).
Thirty days is the common case, not the ceiling. Anthropic keeps inputs and outputs for up to 2 years, and trust and safety classification scores for up to 7 years, when its automated systems flag a chat as violating the Usage Policy. Feedback you submit is kept for 5 years. Features that store data by design, such as Anthropic's Files API, follow their own retention under your control.
OpenAI's stateful features work the same way. Responses stored through the Responses API are kept for at least 30 days, and Assistants objects are "retained indefinitely" unless you delete them, after which they are removed within 30 days (OpenAI, Your data). The retention question is therefore per feature, not per provider.
What Does "Zero Data Retention" Actually Mean?
At both providers, zero data retention (ZDR) means your prompts and responses are not kept after the request is processed. Anthropic states that under ZDR it "does not store customer prompts or responses at rest after the API response is returned" (Anthropic, API and data retention). At OpenAI, ZDR excludes customer content from the abuse-monitoring logs and forces the store parameter to false (OpenAI, Your data).
It is not a setting you switch on. OpenAI's ZDR is "subject to prior approval" and additional requirements, and Anthropic's is arranged through its sales team, enabled per organization. Both apply it only to eligible endpoints and features. Anthropic lists what falls outside: the Console playground, Managed Agents, consumer plans, the Team and Enterprise chat interfaces, the Batch and Files APIs, and code execution.
Both also keep stated exceptions. Even under ZDR, Anthropic may retain data required by law or flagged by its automated trust and safety systems, for up to 2 years. OpenAI states that images flagged for potential child sexual abuse material are retained for manual review and legally required reporting, even in ZDR deployments (OpenAI, 19 August 2026).
What changed in 2026
The most capable models now come with safety retention attached. From 9 June 2026, Anthropic requires 30-day retention for prompts and outputs on its Covered Models, currently including Claude Fable 5 and Mythos 5 and their 5.1 versions, and these models are not available under ZDR unless Anthropic expressly authorizes it. A ZDR organization can enable 30-day retention for one workspace while its others stay at zero. Human review of that data "can occur only through a controlled access path", for example after an automated flag, and access is logged (Anthropic, Covered Models).
OpenAI took a different route. On 19 August 2026 it previewed Private Safety Processing, designed so automated systems can detect misuse across related interactions and return "limited safety signals" without OpenAI personnel seeing the underlying content. It began rolling this out to API customers in phases on 22 September 2026 (OpenAI, Offering Zero Data Retention for frontier models). OpenAI's documentation still reserves the right, with written notice, to make specific models ineligible for ZDR for specific customers, in which case content may be kept in encrypted logs, or in some cases retained and reviewed by people when classifiers flag it (OpenAI, Your data).
Neither approach is unreasonable. Frontier models raise real misuse risks, and both providers are being explicit about the trade. The practical point for a buyer is that "we have ZDR" is now a statement about specific models, endpoints and dates, and it needs re-checking whenever you adopt a new model.
Is "Not Used for Training" the Same as "Not Stored"?
No, and this is the gap most summaries skip. A no-training commitment says nothing about whether a copy of your prompt exists, for how long, or who can open it. On a standard API contract, the honest description is: not trained on, kept for up to 30 days, and readable by a small set of people in stated circumstances.
For most work that is an acceptable trade. For work that carries your intellectual property, pricing, source code, deal terms or client files, it may not be, and the reason is rarely training. It is that the obligation you owe a client or a regulator is usually written in terms of who may hold and process the data, and a 30-day or 2-year copy on someone else's infrastructure is holding it.
That is the honest bridge to running a model yourself. Not that providers are careless, but that only a model you control lets you make all three promises to your own clients without an exception list. We set out how to secure one, and what it costs to operate, in the on-prem versus private cloud versus API comparison.
Which Work Can Go to an API?
Classify the work, not the vendor. Three questions settle most cases, and they come in a fixed order.
The first question is contractual, so it is answered by your agreements, not your preferences. The second is where the three promises matter: if a 30-day copy is acceptable, a standard enterprise API is a sound choice. The third catches the case people most often get wrong, assuming a ZDR agreement covers a feature or model that it does not.
Most companies end up with all three outcomes at once, which is why routing is a design decision rather than a policy memo. The first step of our method, Map your entire company, is where each workload gets classified, and the second, Control the data, is where the routing is built. That is the foundation of our AI business automation work and our custom AI software engagements. We run our own self-hosted model for work that touches internal knowledge, and we built enterprise AI software for Signature Intelligence Systems, for organizational knowledge capture, team management and global logistics, operating in on-prem and SOC 2 compliant cloud infrastructure instances.
Where Is the Evidence Thin?
A reference page should name its limits, so here are ours.
| The claim | Why it is softer than it looks |
|---|---|
| Every cell in the comparison table | Accurate as of the provider pages fetched on 30 September 2026. Both providers changed these terms during 2026, and several of the pages are help articles that are revised without a changelog. Re-read the current page before you sign. |
| Human review on consumer ChatGPT | The consumer data-control pages we checked describe training and deletion but not staff access. We have left that cell unfilled rather than infer it from the business pages. |
| Your own contract | Negotiated agreements, data processing addenda and zero data retention approvals can differ from the public pages. The public pages are the default, not a ceiling or a floor for your deal. |
| Private Safety Processing | OpenAI began a phased rollout to API customers on 22 September 2026. How it behaves at scale, and which customers it reaches first, is not yet something anyone outside OpenAI can verify. |
| Other providers | This comparison covers OpenAI and Anthropic only. Google, Microsoft and the cloud marketplaces that resell these models publish their own terms, and on Amazon Bedrock and Google Cloud the cloud provider, not Anthropic, is the data processor. |
The pattern to take from it: the training commitments are stable and well documented, and the retention and review terms are the ones that move. Build your review process around the columns that change.
Frequently Asked Questions
No, not by default. OpenAI states that it does not train its models on inputs and outputs from ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu or the API unless the customer explicitly opts in, for example through feedback mechanisms. Training is only one promise, though. Enterprise workspace admins set how long conversations are kept, deleted conversations are removed within 30 days unless the law requires otherwise, and authorized OpenAI employees may access conversations to resolve incidents, recover conversations with your permission, or where the law requires it.
No. Anthropic's Commercial Terms state that Anthropic may not train models on Customer Content from its services, and its API documentation states that retained data is never used for model training without your express permission. That covers the Claude API and Claude for Work plans such as Team and Enterprise. Consumer Claude plans are different: Free, Pro and Max users choose whether their chats are used for training, and choosing yes extends retention to up to five years.
It means the provider does not store your prompts and responses after the request is processed, instead of keeping them for up to 30 days for abuse monitoring. At OpenAI and Anthropic it is available only by approval or agreement, only on eligible endpoints and features, and it has stated exceptions: content required by law, content flagged by safety systems, stateful features such as file storage, and, at Anthropic since 9 June 2026, Covered Models that require 30-day retention even for ZDR customers.
By default, up to 30 days for both. OpenAI keeps abuse-monitoring logs, which can contain prompts and responses, for up to 30 days, and Anthropic deletes API inputs and outputs within 30 days. Longer clocks apply in specific cases: Anthropic keeps content flagged as violating its Usage Policy for up to 2 years and the classification scores for up to 7 years, feedback data for 5 years, and objects you store, such as files or OpenAI Assistants threads, stay until you delete them.
No. Training decides whether your data shapes a future model. Retention decides whether a copy of your prompt exists on the provider's systems, and for how long. Human review decides whether a person can read it. Enterprise plans and APIs close the first by default, zero data retention narrows the second, and no commercial tier removes the third entirely, although each provider limits it to narrow, stated cases such as flagged content, incidents and legal requirements.
No. A personal Free, Plus, Pro or Max account runs on consumer terms, whatever the employee pastes into it. On consumer ChatGPT, OpenAI may use content for training unless the user opts out, and on consumer Claude, training depends on the user's own setting. The business terms apply only to work done inside the business workspace or through the API, which is why an enterprise contract does not cover shadow AI.
When a copy on someone else's systems is the problem, not just training. If a contract or regulation limits who may process the data, or a 30-day or flagged-content copy on provider infrastructure is unacceptable and the work cannot run entirely on zero-data-retention-eligible endpoints and models, it belongs on a model you control, self-hosted in your cloud tenant or on-prem. General drafting and research on non-sensitive material can stay on an enterprise API.