1What's Collected — and Where It's Sent
The whole picture, at a glance
Unlike some assistants that split a metadata stream off to a separate analytics vendor, Codex sends everything to one place: OpenAI. When it runs against a ChatGPT account, your prompts, code, and conversation go straight to OpenAI's servers — and by default those responses are stored server-side and, on personal ChatGPT plans, used to train the model. Here is every key item and exactly where it lands.
| Data being collected | What it exposes | Sent to |
|---|---|---|
| ChatGPT account identity | Your account & email tie the session to you — not anonymous | OpenAI |
| Your full prompts & conversations | Every word you type and receive, verbatim | OpenAI |
| Your code & coding work | The actual source you paste, generate, or edit | OpenAI |
| Tool-call arguments & file context | Command arguments, file paths, and content the model acts on | OpenAI |
| Working directory & environment | The folder Codex runs in, model, reasoning effort, token counts | OpenAI |
| Stored responses (store = true, default) | Each response is retained on OpenAI servers so turns can be chained | Stored |
| → Used to train future models (personal ChatGPT plans, default ON) | Your content becomes training data for the next model | OpenAI |
| Local tool execution & output | Runs on your machine; only what's needed for the next turn is sent | Local |
store = true so OpenAI holds prior context server-side and each turn sends only the new message. The practical effect: response objects are cached on OpenAI's servers. This is an optimization, not a requirement — turning it off simply replays the conversation from your own local history each turn. Functionality is identical either way; the only cost is slightly slower turns and higher token use on long conversations. Disabling it removes this cache (see below).
2Training Exposure by Plan
Who is protected, who is not
Whether your content is used to train the model depends entirely on your ChatGPT plan. Personal plans default ON; business plans are protected by contract.
How it breaks down by plan
| Plan | Is your content used for training? |
|---|---|
| Free | On by default Your conversations and code are used to train the model unless you opt out in Data Controls. |
| Plus | On by default Same defaults as Free — a paid consumer plan does not exempt you. Training on unless you switch it off. |
| Pro | On by default Same as Free and Plus — content is used for training unless you opt out. The burden is on you. |
| Team | Protected Workspace data is not used for training by default — a Data Processing Agreement prohibits it. |
| Enterprise | Protected Not used for training by default; contractual DPA. Zero-Data-Retention available for qualified accounts. The recommended tier for sensitive work. |
| Business / API | Protected Not used for training by default under commercial terms — unless you explicitly opt in. |
One caveat about opting out
Turning off model training stops future use for training, but it does not delete past conversations, and OpenAI still retains data for a period (commonly ~30 days) for abuse and safety monitoring. Zero-Data-Retention terms (Enterprise/API) are the only way to eliminate server-side retention entirely.
3Configuring for Maximum Privacy
The switches
Maximum privacy with Codex means addressing both the training setting and the server-side storage default. Here is the checklist we apply for our clients.
A. Turn off model training
Open Settings → Data Controls and switch this off:
- Toggle Improve the model for everyone OFF — stops your conversations and code from being used to train the model.
B. Disable server-side response storage
By default Codex chains turns using store = true, which caches each response on OpenAI's servers. This is a no-loss privacy setting — disabling it sends requests statelessly (the conversation is replayed from local history), so work is completely unaffected:
codex.store_responses = false
The only trade-off: multi-turn conversations, tools, and coding all keep working exactly as before — the sole cost is slightly slower turns and higher token use on long chats, since context is re-sent rather than chained. No functionality is lost.
C. Use a protected plan for real work
- Prefer Team / Enterprise / Business — workspace data is excluded from training by contract, no per-user toggle needed.
- For the highest bar, ask OpenAI about Zero Data Retention (ZDR) on Enterprise / API — the only way to remove server-side retention entirely.
- Never use a personal ChatGPT login for sensitive or client work.
Verification that actually sticks
- Confirm Improve the model for everyone is OFF in ChatGPT Data Controls.
- Set
codex.store_responses = falseto stop server-side response retention. - Use a Team / Enterprise / Business account for anything sensitive.
- Remember: opting out stops future training but not the standard ~30-day safety-retention window — only ZDR removes that.
Bottom line
With Codex, everything goes to one vendor — OpenAI. On a personal ChatGPT plan your prompts and code are stored server-side and used for training by default. Two switches close that: turn off training, and turn off server-side storage.
Kukui IT configures these controls for our managed clients as part of a standard secure-development posture. If you would like us to audit and lock down the AI tooling in your environment, reach out to your account team.
Sources
Primary references · verified July 2026
Facts in this advisory are drawn from OpenAI's own documentation. Settings, defaults, and terms can change — always confirm against the current source and your own account settings.
- OpenAI — Data Controls FAQ — the “improve the model” toggle, retention, and what opt-out does and does not cover.
- OpenAI — What are the Data Controls settings? — where consumer users turn training off.
- OpenAI — Keep history on but disable model training — separating history from training.
- OpenAI — Enterprise privacy — Team / Enterprise / API training exclusion and Zero Data Retention.
- ChatGPT — Data Controls settings — the live settings page.