What the major AI providers actually publish about training on your data, how long they keep it, how much notice you get before a price rise, and who you are really contracting with. Free, no login, and every table carries the date it was checked.
Last reviewed 18 September 2026
Most AI procurement reviews the rate card. Almost none reviews the terms sitting behind it — and those terms differ between providers far more than the prices do. Worse, they change when a vendor edits a web page, with no email and no announcement.
This page collects the published answers to the five questions that come up most often, side by side. Where a section number exists we cite it, so you can open the vendor's own document and check us.
The answer most buyers assume is "no, obviously." It is not universal — two of the six below train by default unless you change a setting.
| Provider | Trains on API data by default | How you obtain zero data retention |
|---|---|---|
| OpenAI | No | Approval, plus additional requirements |
| Anthropic | No | Through sales, enabled per organization |
| Google Gemini API | No on paid tiers. Free tier and AI Studio usage may be human-reviewed | Application per project; no published approval timeline |
| Mistral | Yes — trains by default, with an opt-out. Commercial Terms §4.2 makes the default product-dependent without naming which products, so confirm for your agreement. | Written request with a stated reason, and it can be denied |
| Cohere | Not stated — Cohere documents an opt-out Data Controls toggle but does not publish that it trains by default. Confirm for your agreement. | Enterprise agreement only |
| AWS Bedrock | No | A mode setting — but only for models that permit it |
⚠ Verified 8 September 2026 against each provider's own published documentation. Provider terms change without announcement — re-check before relying on this.
No. At every provider checked, it is a contract term with a scope — and the scope has documented holes the marketing page does not describe.
A security questionnaire asks one question and offers a checkbox. The reality has four separate mechanisms, and answering for one while assuming it covers the others is how a confident answer becomes a wrong one.
| Mechanism | Why it gets missed |
|---|---|
| Training use | Everyone assumes "no by default" is universal across providers. See the table above — it is not. |
| Abuse monitoring | Data held so the vendor can investigate misuse. This survives most zero-retention arrangements. |
| Operational logging | Data a feature needs in order to function. Eligibility is decided per feature, not per account. |
| Sub-processor retention | What the vendor's own vendors keep. The weakest-documented leg across every provider — nobody publishes it. |
Two things that catch people out. First, eligibility is per feature, not per account. Anthropic publishes, for example, that batch jobs retain 29 days, code-execution containers 30 days, and Files API content persists until deleted; Mistral's Privacy Policy names the Agents API and the Fine-Tuning API as stateful exceptions to its thirty-day rule. A customer with a signed arrangement who turns on an agent framework or moves a workload onto batch processing to cut cost has moved data outside it — without signing anything.
Second, some models are excluded entirely. Anthropic designates certain models as Covered Models, which are unavailable under zero data retention unless expressly authorised — even for a customer holding an organization-wide arrangement. Changing a model name in a config file therefore changes the contractual position, with nothing in code review or the API response to signal it.
⚠ Verified 8 September 2026 against each provider's published documentation.
Ask this one separately from everything else. It is the question most likely to produce a number that contradicts the trust page.
| Provider | The exclusion that matters most |
|---|---|
| Anthropic | Flagged sessions held up to two years regardless of arrangement — including under zero data retention and including under a HIPAA business associate agreement. Covered Models are unavailable under ZDR unless expressly authorized by Anthropic. This is legitimate abuse monitoring, not misconduct — but it is a scope difference between the trust page and the contract. |
| OpenAI | Safety-retention override for severe-risk investigations; API inputs and outputs may be securely retained up to 30 days, except for certain endpoints and features listed in OpenAI's platform documentation. |
| Google Gemini API | Search and Maps grounding retain 30 days — Google states “There is no way to disable the storage of this information.” |
| Mistral | Privacy Policy names Agents API and Fine-Tuning API as the stateful exceptions. |
| Cohere | Usage metadata is never covered. |
| AWS Bedrock | AWS states “you always control your retention policy” — an account set to no retention has the request blocked and the model shown unavailable rather than overridden. Review is “carried out by AWS within the AWS boundary — the model provider does not review your content.” |
⚠ Verified 8 September 2026. This is legitimate abuse monitoring, not misconduct — but it is a gap between the trust page and the contract, and it is the reason a healthcare questionnaire answered "zero retention" may have been answered wrongly.
Every procurement reviews the rate card. Almost none reviews the clause governing what the vendor must do before that rate card moves.
| Vendor | Where the clause lives | Stated notice |
|---|---|---|
| OpenAI | Services Agreement §6.6 | 14 days after posting on the Pricing Page |
| Anthropic | Commercial Terms §H.1 | 30 days after posting, or when the customer otherwise receives notice — whichever is earlier |
| Gemini API Additional Terms | 30 days after posting; new paid services can take effect immediately |
⚠ Clause text pulled 14 August 2026. Section numbers cited so you can check them yourself.
Withdrawn 18 September 2026 — DeepSeek. This table previously carried a row describing DeepSeek's price-change notice. We were unable to re-retrieve the source page to confirm it, so the row has been removed rather than left standing. It will return when the page has been read and captured.
Read the trigger carefully: "after they are posted." The clock does not start when the vendor tells you. It starts when the vendor edits a web page. If nobody on your side is watching that page, a full 30-day notice period can pass without a single person at your company seeing it.
The same structure governs model deprecation and rate limits at most providers. Three clause families, one root pattern: the vendor's obligation is discharged by publishing, not by communicating.
Not necessarily the platform. On some platforms the answer is set per model, on the same account.
Withdrawn 18 September 2026. This section previously set out, in a per-model table, which model families on one cloud platform are sold by the platform and which are sold by the model provider, quoting that platform's legal page directly. We could not re-retrieve that page to confirm the quotations, so the table has been removed rather than left standing on an unverified reading. It will return when the page has been read and captured. The general point below stands on its own and names no vendor.
The structural point, which is worth knowing whichever platform you use. Buying a model through a cloud marketplace does not automatically put the cloud provider between you and the model's maker. Depending on how the platform has papered each model, you may be agreeing to the model provider's own terms in full — its indemnification, its liability cap, its governing law — with the platform acting only as the storefront. Switching a workload from one model to another can therefore move your counterparty, and nothing in a code review will flag it.
Two questions are worth holding apart, because buyers routinely merge them. Who is my counterparty can be changed by contract language alone. Where does my data actually go changes only if the technical path changes. A contract can change who you sue. It cannot change what a vendor's systems do.
The question to put to any platform you buy models through: for each model we actually use, who is the seller of record, and whose terms govern — yours, or the model provider's? Ask it per model, and ask for it in writing.
Every entry comes from the vendor's own terms, pricing page or documentation — not from a summary, a review site or a press report. Section numbers cited where they exist.
Each table carries the date it was read. Treat anything older than about sixty days as needing a re-check before a live negotiation.
When something here turns out to be wrong, the correction is published with a date on it. Nothing is quietly edited after the fact.
This page is a snapshot. Terms Watch is the weekly diff — what moved, and what it means for an agreement you have already signed. Free, one email a week.
With your consent this site uses Google Analytics to measure traffic and anonymous, aggregate tool usage — including your score or result band, but never the text you type and never anything identifying you or your company. Google sets a cookie and processes your IP address. See our privacy note.
