You're about to write the sentence in your privacy policy that says what happens to a customer's message after it reaches an LLM API. Most people write "it is not used for training" and stop. That answers one question. How long the vendor keeps the text, who can read it and which features quietly keep more are separate questions, and the answers differ by vendor, plan and feature.
I read each vendor's own data pages on 2026-10-09 and put the answers in one table, with "unknown" wherever the page didn't say. Treat this as a map to check, not a contract. Vendors change these pages, and your contract can override them.
What does each vendor say?
| Vendor and product | Trained on your data by default? | Standard retention of prompts and responses | Zero-retention route | Page and date shown |
|---|---|---|---|---|
| OpenAI API | No. "As of March 1, 2023", API data isn't used to train unless you opt in | Abuse monitoring logs kept "up to 30 days" by default; some endpoints also keep application state (see below) | Zero Data Retention or Modified Abuse Monitoring, by prior approval | OpenAI "Data controls in the OpenAI platform" guide, no date shown |
| Anthropic API | No, by default. Feedback you submit, or data you choose to share, may be used | Privacy centre: deleted from the backend "within 30 days of receipt or generation", with exceptions. API docs: content "not retained by default" except for models that require 30-day retention | Zero data retention by agreement, per organisation | Privacy centre articles dated 1 July 2026 and 18 August 2026; API docs page, no date shown |
| Google Gemini API, paid (billing-enabled project) | No. Prompts and responses aren't used to improve products | Logged "for a limited period of time" for abuse and legal reasons. Duration unknown on the terms page | Unknown on the terms page | Gemini API Additional Terms of Service, "Effective March 23, 2026" |
| Google Gemini API, unpaid quota and AI Studio without billing | Yes. Used to improve products; human reviewers may read inputs and outputs | Unknown | Not offered on that page. In the EEA, Switzerland and the UK the paid-service terms apply even to unpaid use | Same terms page |
| Google Cloud enterprise platform (the "Gemini Enterprise Agent Platform" page) | No training without your permission | Abuse-monitoring prompt logging may apply; 24-hour in-memory cache by default; some features store longer | Request an abuse-monitoring exception; avoid listed features | Google Cloud "zero data retention" page, no date shown |
| Azure OpenAI / Foundry models | No. Not used to train foundation models without your permission | Models are stateless. A flagged sample may be stored for human review. Period unknown on the pages I read | "Modified abuse monitoring" by approval | Microsoft Learn "Data, privacy and security" page, no date shown |
| AWS Bedrock | Not stated on the page I read | Does not store inputs or outputs by default; listed models retain traffic up to 30 days for abuse detection | Described as the default. For the listed OpenAI models, eligible customers may request full ZDR through their AWS account team; for the listed Anthropic models the page mentions a safeguards programme with ZDR only through 31 December 2026 | Bedrock "abuse detection" page, no date shown |
| Mistral API (Studio) | Free mode: may be used, opt out available. Pay-as-you-go: "right to opt out"; default not clear | Unknown | Page links to a separate ZDR article I did not read | Mistral Help Center, "August 31, 2026" |
Not checked at all: xAI, Cohere, Groq, DeepSeek, Perplexity's API, Meta's hosted Llama services, and every consumer chat app. For consumer settings, see what not to paste into ChatGPT.
How I read the pages: I fetched each page directly and read the extracted text, so the quotes in the table are from the page text I saw. I read Google's and Microsoft's pages as plain text without their navigation, and some had long lines my tool truncated. Re-read the exact wording on the live page before you quote it.
What does each row hide?
OpenAI. The headline is "not used for training" and "up to 30 days" for abuse logs. The detail is in the endpoint table. On the page I read, Chat Completions and Responses list 30 days of abuse monitoring and no application state retention, with exceptions, while the Responses API "has a 30 day Application State retention period by default, or when the store parameter is set to true". The /v1/conversations objects show "Until deleted". Assistants objects are deleted 30 days after you delete them, and objects you never delete are "retained indefinitely". Prompt caching may hold encrypted tensors for up to 24 hours. Zero Data Retention forces store to false. Image and file inputs are scanned for CSAM and retained for review if flagged, even under Zero Data Retention.
Anthropic. The two pages I read describe retention differently. The privacy centre says API inputs and outputs are automatically deleted within 30 days. The API docs say conversation content isn't retained by default, with models designated "Covered Models" requiring 30-day retention. Both mention exceptions for usage-policy enforcement. On the privacy centre page, flagged content can be kept for up to 2 years and trust and safety classification scores for up to 7 years. Feedback data you submit is kept for 5 years. The Files API keeps files until you delete them or they expire, and batch processing has 29-day retention. Code execution containers retain data up to 30 days. Because the two pages differ in wording, read your contract and ask your account rep which applies to you.
Google. The biggest trap is the free tier. Using the Gemini API without a billing-enabled project means your content can be used to improve products and read by reviewers, and the page says "Do not submit sensitive, confidential, or personal information to the Unpaid Services." On the enterprise platform page, "Grounding with Google Search" keeps logs for up to three days with no way to disable them, "Grounding with Google Maps" stores prompts and output for thirty days, and the Interactions API with store unset defaults to true.
Azure. The Responses API and Assistants threads store data in your own Azure tenant, in the same geography, and the page says you can delete it at any time. I didn't see an automatic expiry on that page. A separate abuse-monitoring store holds flagged samples for human review, and the page says it's logically separated by customer resource. I did not find a retention period for that store on the pages I read, so the table says unknown.
AWS Bedrock. Plainest of the lot: no storage of inputs or outputs by default. But the abuse page lists specific models whose traffic (all traffic for the listed Anthropic models, classifier-flagged traffic for the listed OpenAI models) is kept up to 30 days, and says retained data is stored and processed by AWS and not shared with the third-party model provider. It also notes that with cross-region inference the retained data sits in the region where the request ran.
Mistral. The page I read covers training only. For Studio's free mode the company may train on your data unless you opt out. For pay-as-you-go the wording is that customers "retain full control" and "have the right to opt out at any time", which doesn't tell me the default. Retention isn't on that page.
How should you use this table?
Treat it as a pre-flight check before you pick a vendor or write a privacy notice.
- Find your exact plan. "OpenAI" or "Google" is not one policy. Unpaid versus paid Gemini, or API versus consumer chat, changes the answer completely.
- List the features you use. File uploads, stored responses, batch jobs, caching, code execution and web grounding each have their own retention. A stateless chat call is the best case.
- Write the retention into your vendor register, with the date you read it. Copy this row format (untested template):
- vendor: <name>
product_and_plan: <API, paid, region>
page_url: <the vendor page>
date_read: 2026-10-09
training_default: <yes/no/unknown, quote the line>
standard_retention: <number and condition, or unknown>
zero_retention: <available? approval needed?>
features_used_that_store_more: [files, stored responses, batch]
contract_or_dpa_signed: <yes/no>
- Reduce what you send. If the text never contained a phone number, nobody needs to retain it. See PII redaction before LLM calls.
- Turn off storage you don't need. Where a vendor offers a
storeswitch or a cache setting, check what its default is for your endpoint and set it deliberately. - Re-check every quarter. Several of these pages carry dates from the last few months, so they do change.
- If nothing here is acceptable, keep the data in-house. A self-hosted model removes the vendor leg. For sizing, see a private ChatGPT with Ollama and Open WebUI.
If your users are in India, vendor retention feeds straight into your own duties: you need to know where personal data sits to honour erasure and breach duties. The DPDP checklist maps those duties to tasks. None of this is legal advice, and a lawyer should review the wording you publish.
What I couldn't verify
- Gemini paid-tier log duration. The terms page says "limited period" and gives no number.
- Azure abuse-monitoring retention period. Not found on the pages I read.
- Mistral's default and retention. Not on the page I read, and I didn't open its ZDR article.
- Whether your contract differs. Enterprise agreements, DPAs and regional terms can override public pages.
- Anything after 2026-10-09. Pages change.
- Vendors not listed.
