The change

OpenAI updated its Usage Policies on October 21, 2026, with an effective date of November 20, 2026. The changelog describes the update as clarifying “restrictions on using OpenAI models and outputs to train other models,” along with protections for children and teens, requirements for high-stakes uses, and approval rules for national security, law enforcement, and model-access intermediaries.

The biggest addition is a new section titled “Training and developing models.” OpenAI’s terms already said you could not use output to build competing models. The new section spells out, step by step, what counts, and it reaches work that many teams would not call training at all.

What changed

• Training rival models. “You must not use OpenAI models or their outputs to generate, prepare, select, or supply training material or signals for models that are intended as alternatives to or substitutes for OpenAI models or services.”

• Reasoning extraction. “You must not attempt to extract protected reasoning through unauthorized access, reverse engineering, or circumvention of safeguards.”

• Licensed advice. The old rule banned “provision of tailored advice that requires a license.” The new rule targets “building or offering AI tools that provide others with tailored advice that requires a license,” and adds: “You also may not state or imply that the AI itself holds professional credentials or authorization.”

• High-stakes uses. The human review rule moved from “high-stakes decisions” to “high-stakes actions,” including “actions, decisions, transactions, submissions, and other steps.”

• Model access. “If you operate a model-access intermediary platform or other service that enables third parties to directly access OpenAI models, you must obtain our prior review and approval.”

From “compete” to every step of training

OpenAI’s Terms of Use already list “Use Output to develop models that compete with OpenAI” among things you may not do, and its Services Agreement for business customers has a similar clause. Those lines are short and leave open what “develop” means. The new Usage Policies section answers that. It names three kinds of work:

• “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.”

• “Selecting, filtering, or validating training data, tasks, or trajectories.”

The second item is the one most likely to surprise teams. Using a GPT model to grade or rank answers is common practice, and if those grades feed the training of a model meant to replace OpenAI’s, the policy now names it directly.

The section also closes the indirect routes. “These restrictions apply whether you perform 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.” When the model is meant as a substitute for OpenAI's, a data labeling vendor that uses ChatGPT for a client, or a team that later hands its evaluation results to someone training that model, falls inside that wording.

What is still allowed

The policy is explicit that normal work continues: “You may use OpenAI models for research, software development, model evaluation and benchmarking, and product development, within these boundaries.” The restrictions also “do not prohibit uses expressly permitted by your applicable Agreement with OpenAI, including any Permitted Exception under the Services Agreement.”

In the Services Agreement, a Permitted Exception covers two cases: models “primarily intended to categorize, classify, or organize data (e.g., embeddings or classifiers), if these models are not distributed or made commercially available to third parties,” and fine-tuning or customizing models that OpenAI itself provides. Everything else that trains a substitute model is now covered in detail.

What did not change

• The privacy rules, including the bans on facial recognition databases without consent and on social scoring, are unchanged.

• The bans on fraud, scams, impersonation, and academic dishonesty are unchanged.

• The Terms of Use still carry the short “compete with OpenAI” clause alongside the new, detailed section.

• Apps that simply use OpenAI models to power their own features do not need the new approval: the rule “does not apply solely because an application integrates OpenAI models to power its features or functionality.”

• At the time of publication, the previous version dated October 29, 2025, applies until November 20, 2026.

Why this matters

AI providers are writing more precise rules about how their models may be used, and those rules now reach ordinary engineering work. Anthropic did something similar this month when it let AI approve a claim without a human, but not deny one. OpenAI has also been widening what users answer for, as when its terms for dots made users responsible for what they do and when its app terms began covering ChatGPT calling your app on its own.

For a SaaS company, the training rule matters most when several AI vendors sit in the same stack. If a team uses OpenAI to grade outputs, generate examples, or clean data for an in-house or open model, it needs to know which side of the line that work falls on. That is part of monitoring the AI vendors in your stack, not only the subprocessor lists.

Potential impact

For a company that uses OpenAI models, the update raises five practical questions:

• Do any of your pipelines use OpenAI outputs as training data, labels, scores, or rankings for another model?

• Is that model meant to replace an OpenAI model or service in your product?

• Do your data or labeling vendors use ChatGPT or the OpenAI API on your training projects?

• Does your product give others direct access to OpenAI models, which now needs OpenAI’s approval?

• If your product gives advice in a licensed field, does a licensed professional take part, and does the product avoid presenting the AI as credentialed?

How Venpo detected it

Venpo monitors OpenAI’s legal documents and flagged the new section when it appeared on the Usage Policies page, with every inserted and deleted line marked. The record is on the OpenAI change page. Every quote in this article was checked against the redline and OpenAI’s live pages.

Business outcome

Teams that track OpenAI’s policies get a month before the November 20 effective date to map where OpenAI outputs flow, talk to their data vendors, and move any affected work. That window is what continuous monitoring of vendor legal changes is for.

Key takeaway

From November 20, 2026, OpenAI’s Usage Policies name labels, scores, rankings, and data selection as training work you cannot do with its outputs for a substitute model, even for someone else, so any pipeline that feeds OpenAI output into another model needs a check now.