title: AI vendors — free vs paid use, API credits, and data treatment status: research note (input to the capacity model and MCL) author: Claude (chat, MMDEP-PM role) date: 2026-10-11 question (Atul, 11 Oct 09:41): if a free user buys API credits, how are they treated across the vendor's tools? Also: free accounts and free resources for routine work only, with no confidential knowledge in bulk; "semi-paid" accounts (for example USD 10 of Gemini credit) used in emergencies or where there is ROI
AI vendors: free vs paid use, API credits, and data treatment¶
1. The short answer¶
Buying API credits puts your API usage under the vendor's business terms (not used for training; 30-day retention at Anthropic). It does not change how the consumer apps treat you. The chat apps (claude.ai, ChatGPT) keep their consumer rules. Their training settings decide what happens to your chats.
2. By vendor¶
| Consumer app (chat) | API credits (pay as you go) | Coding tools | |
|---|---|---|---|
| Anthropic | Free, Pro, Max: consumer terms. Chats train future models only if the setting is on (claude.ai → Settings → Data privacy controls). Retention: 5 years if on, 30 days if off | Commercial terms: not used for training; 30-day retention | Claude Code with a subscription login → consumer terms (the training setting applies). With an API key → commercial terms. Cloud sessions follow the account type |
| OpenAI | Free, Plus, Pro: training on by default; turn off under Settings → Data Controls → "Improve the model for everyone" | API platform: not trained on by default | Codex signed in with ChatGPT → ChatGPT's data controls apply. Codex with an API key → API platform |
| Google (Gemini) | Gemini app: consumer terms (separate) | Once at least one API project has billing enabled, prompts are handled under paid terms (not used to improve Google's products). Unpaid API: content used for improvement and read by human reviewers; "Do not submit sensitive, confidential, or personal information" | Gemini CLI with a personal Google login is not the paid API path. Use a key from a billed project for paid terms |
3. Terms that limit how we use the consumer apps¶
| Vendor | Clause (quoted) | What it means for MMDEP |
|---|---|---|
| Anthropic consumer terms | No access "through automated or non-human means, whether through a bot, script, or otherwise", except via an API key or where Anthropic explicitly permits it. "You also may not make your Account available to anyone else" | Scripted use of Claude goes through Claude Code (whose setup-token is documented for scripts and CI) or an API key, never by automating claude.ai. One person per account |
| OpenAI Terms of Use | You may not "automatically or programmatically extract data or Output"; nor "circumvent any rate limits or restrictions"; nor "share your account credentials or make your account available to anyone else" | Correction to 10 Oct: relaying prompts into ChatGPT's web app through Claude in Chrome and reading the replies back would be programmatic extraction of Output. Withdrawn. Agent-to-agent work with ChatGPT goes through issues (ChatGPT's own connector and scheduled tasks) or the OpenAI API |
| Google Gemini API terms | Free tier: content used for improvement and human review | Free Gemini only for public, synthetic, or small non-confidential pieces |
4. What this means for "semi-paid" capacity¶
| Kind of capacity | Examples | Allowed content |
|---|---|---|
| Paid terms (business or API) | Claude API credits; OpenAI API credits (Codex with an API key); a billed Gemini project | Internal and private repository content (still no client data, spec §5B) |
| Consumer, training off | Claude Pro/Max with the training setting off; ChatGPT Plus with "Improve the model" off | Internal work content; no client data |
| Free or consumer, training on | Free ChatGPT accounts left on default; the unpaid Gemini API | Routine work only: small issues, partial information, public or synthetic data (Atul's rule of 11 Oct) |
5. A useful consequence: Claude API credits protect Atul's workplace¶
Unattended Claude jobs (MCL's daily judge S4, scheduled reviews) can run on Claude API credits instead of the subscription. API credits are a separate allowance, so those jobs no longer consume the 5-hour and weekly limits Atul and the PM chat rely on. They also fall under commercial terms (no training). The cost is per token, small for a daily summary with Sonnet or Haiku. This is the same "semi-paid" pattern as the Gemini USD 10 credit: a small prepaid balance, used where it protects the workplace or has a clear return.
6. Actions¶
- Atul: check the Claude training setting (claude.ai → Settings → Data privacy controls), and switch off "Improve the model for everyone" in every ChatGPT account used for work.
- Atul: decide on small prepaid API balances: Gemini (USD 10, agreed); Claude API (for unattended MCL runs); OpenAI API (only if Codex runs unattended).
- MCL design: S4 and scheduled reviews run on Claude API credits when available; the subscription stays for Atul and the PM chat.
- Capacity model: classify each resource by the three kinds in §4, then apply the risk factors: terms, data class, limit volatility, account loss, and dependence on Atul's laptop.
7. Update (Atul, 11 Oct 10:05)¶
- Browser automation for ChatGPT is input-only: push Atul's prompts into the chat window; never read the replies, since all results come back through repositories. That avoids the "extract Output" clause. Feasibility is to be examined later, against the Terms in force then (A-12). The design doesn't depend on it.
- Claude Code: with scheduled tasks,
/loopand cloud sessions, MMDEP doesn't need to call the CLI itself. The vendors' agents work connected to repositories, under the vendors' own guidelines. - The aim is protecting proprietary information and trade secrets, not withholding everything from models. Sharing ordinary work content is acceptable and helps the models improve.
- This is capacity planning, not workarounds. Accounts move to business and enterprise plans as agents prove their ROI. Budgets are decided at the end of the discussion.