OpenAI has published a revision to its Usage Policies that is scheduled to take effect on November 20, 2026. The company says the October 21 update clarifies rules for using its models and outputs to develop other AI models, and adds requirements for some services that give people direct access to OpenAI models. The date matters: as of October 25, the new version has been announced but is not yet effective. OpenAI’s policy page lists the changes and their effective date, while the company’s announcement explains its approach to applying safeguards across services.
For most people using AI to draft an email, summarize notes, or write ordinary software, this is not a notice to stop everyday use. The part developers should examine closely is narrower: what happens when OpenAI’s models or outputs become material for training a model designed to compete with or substitute for OpenAI’s services.
What the model-training rule says
The updated policy says users 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. It gives examples that go beyond collecting a pile of generated text. The listed activities include making training data, tasks, rubrics, or reinforcement-learning environments; supplying feedback such as labels, scores, rankings, preferences, rewards, or corrections; and selecting, filtering, or validating training data, tasks, or trajectories. OpenAI’s updated policy describes the full scope.
In plain language, the rule is about the purpose and role of the work. A team should not assume it is outside the restriction merely because it calls the work “evaluation,” hires a contractor to do it, or does not personally run the final training job. OpenAI says the restriction also covers work done for someone else or through another person, and cases where outputs are later supplied for the training purpose.
That is different from a blanket ban on research or model development. The same policy says OpenAI models may be used for research, software development, model evaluation and benchmarking, and product development within the stated boundaries. It also recognizes exceptions expressly permitted by an applicable agreement. If an organization’s contract includes relevant permission, the agreement—not a generalized summary—needs to be checked. OpenAI’s Services Agreement is linked from the policy as one place where permitted exceptions may be stated.
There is an important timing caveat for API customers. OpenAI’s current Services Agreement already says that, except for a permitted exception, a customer may not use Output to develop AI models that compete with OpenAI’s products and services. That means the November 20 date should not be read as a temporary green light for a workflow that is already covered by a customer’s agreement. Contract terms can differ, so organizations should check the version that applies to them.
Why AI gateways should review their setup
The revision also states that an organization operating a model-access intermediary must obtain OpenAI’s prior review and approval if it enables third parties to directly access OpenAI models. OpenAI’s announcement specifically points to model routers and AI gateways as examples. The policy makes an important distinction: an app does not become an intermediary solely because it uses an OpenAI model to power its own features. The announcement and the policy text describe these boundaries.
That means the architecture matters. A product that calls a model behind the scenes to deliver its own service is not automatically the same as a platform that lets customers directly route their own requests to an OpenAI model. Businesses building routers, gateways, or model-access platforms should check how users reach the model and ask OpenAI about approval before the new effective date if their service may fit the intermediary description.
Other clarifications in the update
OpenAI says the revision also clarifies protections for children and teens, safeguards for high-stakes uses, and review requirements related to national-security and law-enforcement uses. The current policy describes human review for high-stakes actions that could materially affect rights, safety, access to essential services, or other consequential interests. It also says certain national-security or law-enforcement uses and some model-access services require specific approval. These are not simply new labels for every AI project; readers should check the exact conditions in the policy rather than infer that all uses in a broad sector are prohibited.
Comparing the current text with the previous version, effective October 29, 2025, helps show why developers should read the update directly. The older Usage Policy page did not have the new detailed model-training section. It already contained restrictions on high-stakes decisions without human review and on national-security or intelligence purposes without review and approval. The separate Services Agreement also had a competition-related output restriction for its customers. The comparison should not be treated as a full account of every contract term or service-specific rule.
What to do before November 20
If you build or operate an AI product, start by mapping where OpenAI-generated material goes. Ask whether any model output, feedback, label, score, task, rubric, or evaluation result is being used to train or shape a model intended as an alternative to OpenAI’s services. If so, review the exact policy and your applicable agreement before that workflow continues under the new rules.
If your service provides customers with direct access to OpenAI models, document how requests are routed and whether your platform functions as a model-access intermediary. Check the prior-approval language and contact OpenAI if the boundary is unclear. Teams using AI for consequential decisions should also confirm that required human review is built into the workflow rather than added only after an automated decision has already taken effect.
For ordinary users, the practical step is simple: keep using the service for normal tasks in line with its policies, and do not assume this update changes every use of AI-generated text. For developers, the useful deadline is November 20—not because every project must stop, but because model-training pipelines and direct-access services deserve a careful review before the new rules apply. This is a summary of OpenAI’s published policy, not legal advice; organizations with borderline or contract-dependent cases should consult the governing agreement and qualified counsel.
Sources
- OpenAI Usage Policies — current policy text and October 21, 2026 changelog.
- Broad access through consistent standards — OpenAI announcement dated October 21, 2026.
- OpenAI Usage Policies, previous version — version effective October 29, 2025.
- OpenAI Services Agreement — customer restriction related to developing competing models, subject to permitted exceptions.
Leave a Reply