It depends on your tier, not the brand. On a personal Claude plan your data trains future models when that setting is on, and is kept five years. On Gemini, human reviewers read a subset of chats. On the OpenAI API, nothing trains the models unless you opt in. A model on your own machine sends nothing.
What each provider says about its own tiers
| Tier | Used to train models | Retention | Human review |
|---|---|---|---|
| Claude Free, Pro, Max — setting on | Yes | 5 years | Not stated |
| Claude Free, Pro, Max — setting off | No | 30 days | Not stated |
| Claude Team, Enterprise, API | No, unless the customer opts in | 30 days standard | Not stated |
| Gemini Apps — Keep Activity on | Yes, to improve Google services | Saved to your Activity | A subset, by human reviewers |
| Gemini Apps — Keep Activity off | No | 72 hours | Still possible, for safety |
| OpenAI API | No, unless you opt in | Abuse logs up to 30 days | Not stated |
| A model on your own machine | Nothing is sent | Nothing to retain | None |
The sentences those rows came from
- Anthropic. "We will train new models using data from Free, Pro, and Max accounts when this setting is on." Retention is "5-year retention period" for users who allow it, "30-day retention period" for users who do not, and 30 days as standard for Team, Enterprise and API. Commercial users "maintain existing policies: Anthropic does not train generative models using code or prompts sent to Claude Code under commercial terms, unless the customer has chosen to provide their data". (data usage)
- Google. "A subset of chats are reviewed by human reviewers (including Google's trained service providers) to help improve Google services", and the notice asks you directly: "please don't enter confidential information that you wouldn't want a reviewer to see or Google to use to improve our services." With Keep Activity off, "your future chats won't appear in your Activity, and won't be used to train our AI models" — but they "are retained with your account for 72 hours". (Gemini Apps privacy notice)
- OpenAI, API only. "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)", with "abuse monitoring logs … retained for up to 30 days, unless longer retention is required by law", and a zero data retention option that "excludes customer content from abuse monitoring logs". (your data)
The row that is missing, and why
The ChatGPT consumer tier is not in that table. On the day this was written, OpenAI's own policy pages and help centre would not open to us, so we could not read their wording, and the alternative was to repeat what two competing products say about a third party's terms. That is not evidence. If you are about to upload something you would mind losing control of, read OpenAI's own current pages first rather than ours.
Which is the general rule here, and the reason this post is a table of quotations. Every page that ranks for this question is published by somebody selling either the upload or the alternative to it. Including us.
Four things the tiers do not change
- Deleting a chat is not the same as the data being gone. Anthropic's documentation is specific that deleting a session "permanently removes the session's event data" — which is a stronger statement than most, and still says nothing about a backup window.
- A local copy may exist too. Claude Code stores transcripts on your own
disk "in plaintext under
~/.claude/projects/for 30 days by default". - Running the tool locally does not make the model local. A CLI on your Mac still sends the prompt, and a file it reads is part of the prompt.
- Redaction is work you will do badly. Account numbers in a statement, names in a transcript, faces in a photo. The reason "redact it first" is unsatisfying advice is that the file was useful because of the detail you were asked to remove.
The version with no row to read
Every cell above exists because the file went somewhere. A model running on your own machine has no tier, no retention period and no setting to check, and that is the only answer in this table that cannot change when a policy page is edited.
Nico is built on that assumption. No account, no upload: no request it makes carries your files, your voice or anything they produce. The prompts you type go to the Claude Code or Codex on your own Mac, under your own subscription, or to a local model through Ollama — and three things ever leave the machine: a model download, a monthly licence check, and a daily update check carrying nothing.
← All notes