What the AI Vendors Say They Keep
Ask a vendor how long it keeps your prompts and you will usually get the number 30 days. All three big AI platforms publish that figure, and on each page it answers a different question: deletion on one, abuse logging on another, grounding storage on a third. Meanwhile the longest clock any of them publishes runs to seven years. This note reads the three retention pages side by side and separates the clocks. Figures read from vendor documentation on October 12, 2026. Nothing here is legal advice, and none of it replaces the contract your own account sits under.
Free and paid are different products for this purpose

Google draws the sharpest line. Its Gemini API terms say that for unpaid services, submitted content is used to provide, improve and develop Google products and services and machine learning technologies.
The same terms state that human reviewers may read, annotate and process API input and output, and that Google disconnects this data from the Google Account, API key and Cloud project before reviewers see it.
For paid services the terms say Google does not use your prompts, including system instructions, cached content and files, or your responses to improve its products. It logs prompts and responses for a limited period solely for detecting and preventing violations.

OpenAI states the paid position for its API flatly: data sent to the OpenAI API is not used to train or improve OpenAI models unless you explicitly opt in, a position its page dates to March 1, 2023.
Anthropic's retention article covers its consumer products, Free, Pro, Max and Claude Code, and says commercial products including Claude for Work and the API are governed separately. Reading the consumer page and assuming it covers a team plan is a mistake.
Thirty days appears three times and means three things
- Anthropic: deleted conversations are removed from back-end storage systems within 30 days
- OpenAI API: abuse monitoring logs are retained up to 30 days, unless longer retention is required by law
- Google: grounding with Google Search stores prompts, contextual information and output for thirty days; grounding with Google Maps does the same
- Anthropic batch processing: request and response data is stored for up to 29 days after batch creation
- OpenAI: audio outputs in the chat and responses endpoints are stored for one hour to enable multi-turn conversations
Only the first of those is a deletion promise. The second is a safety log you cannot see. The third applies to one feature and starts when you use it, not when you delete anything.
The 29 day figure is a reminder that features have their own clocks. We covered what that means for a results pipeline in our note on batch pricing.
The longest clocks belong to the safety systems
Anthropic's consumer retention article is unusually specific, and the numbers are much larger than 30 days. They apply to categories of data most users never think about.
- Chats and coding sessions used to improve Claude, when a user has opted in: retained in de-identified form for up to five years
- Feedback submitted through thumbs up, thumbs down or bug reports: retained for five years
- Inputs and outputs in chats flagged as policy violations: retained for up to two years
- Trust and safety classification scores: retained for up to seven years
Seven years for classification scores is the longest published retention figure we found on any of the three platforms. It is not your conversation text, but it is a record about your account.
None of this is hidden. It is on a support page, written in plain language, and almost nobody opens it before signing up.
Zero data retention is a list of endpoints, not a switch
OpenAI's data page names the endpoints that qualify for zero data retention: chat completions, responses, embeddings, audio transcriptions and audio translations. Assistants and threads do not qualify.
There is a side effect worth planning around. With zero data retention enabled, the store parameter is always treated as false even if a request sets it to true, so any feature that depends on stored state stops working.
The same page notes that chat completions used with prompt caching may retain data for up to 24 hours. Caching buys the discount described in our caching note and spends a day of retention to do it.
That trade is the clearest example of why the two questions cannot be answered separately. Cost decisions and retention decisions are the same decision.
Seat plans publish contracts, not day counts
Buyers of ChatGPT Business, Claude Team and Microsoft 365 Copilot Business are reading a different kind of document. The commitments are contractual rather than numeric.
Microsoft's Copilot documentation is the clearest example. It says prompts and responses are protected by Enterprise Data Protection under the same contractual terms customers trust for email in Exchange and files in SharePoint.
On location, the same page says Copilot routes model calls to the closest data centres but may call other regions at high utilisation, while European Union traffic stays within the EU Data Boundary.
That is a commitment about jurisdiction, not a number of days. It answers a compliance question and leaves the deletion question to the service agreement.
Four questions worth sending before you sign
- Which document applies to my plan: the consumer retention page or the commercial terms?
- How long is abuse or safety logging kept, and can anyone on my side see it?
- Does any feature I plan to use carry its own retention window, such as caching, batch or grounding?
- If I enable zero data retention, which features stop working?
Every one of those has a published answer on at least one of the three platforms, which makes a vendor that will not answer in writing the more interesting data point.
Bottom line
Do not ask how long a vendor keeps your data. Ask which clock, because there are at least four: deletion, safety logging, opted-in training and feature storage. They run from one hour to seven years.
For a plan-level view of the same three vendors, see our guide to AI assistant plans for small teams and the ChatGPT Business against Claude Team comparison. Admin controls and connected apps carry their own exposure, covered in an earlier brief.
Sources
- Your data, OpenAI API documentation - checked October 12, 2026
- How long do you store my data, Anthropic privacy centre - checked October 12, 2026
- Gemini API additional terms of service - checked October 12, 2026
- What is Microsoft Copilot, Microsoft Learn - checked October 12, 2026
- Batch processing, Claude Platform documentation - checked October 12, 2026