Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

OpenAI's training ban lands November 20

Written November 9, 2026

OpenAI's usage-policy update (Oct 21, effective Nov 20) bans using outputs to train competing models. What's banned, the exception, and the Anthropic contrast.

What the rule says

OpenAI's Usage Policies, updated October 21, 2026, add a section called "Training and developing models." From November 20, 2026, you must not use OpenAI models or their outputs to generate, prepare, select, or supply training material or signals for models intended as alternatives to or substitutes for OpenAI models or services. The policy's reasoning is short: training another model on OpenAI outputs can transfer capabilities without carrying the safeguards over.

The ban reaches further than a full training run. It covers creating or preparing training data, training tasks, rubrics, or reinforcement-learning environments; providing feedback used in training, including labels, scores, rankings, preferences, rewards, or corrections; and selecting, filtering, or validating training data, tasks, or trajectories. It also bans attempting to extract protected reasoning through unauthorized access, reverse engineering, or circumvention of safeguards.

Two scope lines matter for teams. The restrictions apply whether you do the work yourself, for another party, or through someone acting on your behalf, even if you do not train the resulting model. They also apply when outputs from product use or evaluation are later used or supplied for those training purposes — so a labelling pipeline fed by eval outputs is covered, not just one fed by the API directly.

What is still allowed

The same section says you may use OpenAI models for research, software development, model evaluation and benchmarking, and product development, within these boundaries. Evaluation stays allowed; feeding the outputs of that evaluation into a substitute model does not. The policy also says the restrictions do not prohibit uses expressly permitted by your applicable agreement with OpenAI, including any Permitted Exception under the Services Agreement.

The exception, word for word

The Permitted Exception sits in the Services Agreement for business and developer customers (effective January 1, 2026). It covers using output to (a) develop models primarily intended to categorize, classify, or organize data — embeddings or classifiers, for example — if those models are not distributed or made commercially available to third parties; and (b) fine-tune or customize models provided as part of OpenAI's fine-tuning or other services set out on the Pricing Page. That is the whole exception as written: internal classifiers that stay internal, and fine-tuning OpenAI's own models. It does not cover training a competing model, and it does not cover a classifier you sell or give to third parties.

The Anthropic contrast

Anthropic's rule is stricter, and it lands first. Its Usage Policy, effective November 12, 2026, bans using outputs to train an AI model — naming "model scraping" and "model distillation" — without prior authorization from Anthropic. Where OpenAI's ban is scoped to models intended as alternatives or substitutes, Anthropic's covers any AI model unless Anthropic has authorized it in advance. A team that trains on both providers' outputs faces the stricter rule eight days earlier, on November 12.

One more change in the same update

The October 21 update also adds an approval rule for model-access intermediaries: if you operate a platform or service that enables third parties to directly access OpenAI models, you must obtain OpenAI's prior review and approval. An application that simply integrates OpenAI models to power its own features is not covered by that rule. The update also sets out approval requirements for national security and law enforcement uses.

Verdict

If any pipeline turns OpenAI outputs into training signal for a rival model, stop it before November 20. The rule covers labels, scores, and rankings — not just training runs — and it covers work done for someone else and outputs that arrive through evaluation. The only safe harbours in the text are internal classifiers that stay internal and fine-tuning OpenAI's own models.

Steps

  1. List every place OpenAI outputs go. API calls, ChatGPT sessions, eval harnesses, labelling queues. Mark each destination: training data, a rubric, a ranking, a filter, or a model.
  2. Ask whether the target model is an alternative or substitute. If it is, the pipeline is banned from November 20, even if the training itself happens elsewhere and even if someone else runs it.
  3. Check the exception before relying on it. Confirm which agreement covers your account, that a classifier is not distributed or sold to third parties, and that any fine-tuning is of OpenAI's own models through OpenAI's services.
  4. Cut eval outputs out of training sets. If an eval harness writes labels or scores that later feed a training run, break that link or stop the feed.
  5. Apply the stricter date if you also use Claude. Anthropic's ban takes effect November 12 and covers any model without prior authorization, so a shared pipeline needs to satisfy it first.

What can go wrong

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: GPT-5.5-Cyber preview ends February 1, 2027 · DHS DeepSeek order: inventory your stack · Anthropic's no-distillation rule lands November 12 · Goldman Sachs limits Claude Code to engineers