OpenAI today began requiring operators of platforms that give third parties direct access to its models to obtain the company's prior review and approval, under a rewritten version of its Usage Policies that also adds a section barring the use of OpenAI models and outputs to prepare training material for rival AI systems. The new text, which carries an effective date of November 20, 2026, replaces the version that had applied since October 29, 2025. OpenAI announced the changes in a changelog entry dated October 21, 2026, which gave customers 30 days before the rules took hold.

In Short

From today, a platform that lets other people or companies tap directly into OpenAI's models, such as an AI model router or gateway, needs OpenAI's review and approval first, while an app that only uses the models to run its own features does not. The same rewrite says OpenAI models and their outputs may not be used to build training data for AI systems meant to replace OpenAI's, rewrites the child safety rules, and widens the human review rule from automated decisions to automated actions. For anyone buying AI tools, the line between a vendor that passes on model access and a vendor that builds features on top of a model now carries a different set of obligations.

What OpenAI says changed

The changelog entry is the closest thing to an official summary. "We've updated our Usage Policies to clarify restrictions on using OpenAI models and outputs to train other models," it reads. "We've also clarified protections for children and teens and requirements for high-stakes uses. The update also incorporates our existing national security principles and sets out approval requirements for national security and law enforcement uses, as well as for model-access intermediaries. These changes are effective November 20, 2026."

A help center notice titled "We're updating our Usage Policies" lists the same four themes. According to that notice, the previous version stayed in force until today, and existing agreements and other applicable terms, including any express permissions granted in them, continue to apply. It describes the training language as a clarification of restrictions that already existed rather than a new prohibition, and it describes the intermediary rule as an added approval requirement.

The word "clarify" carries a lot of weight in that summary. The policy body grew from about 830 words in the October 2025 version to about 1,400 words in the new one, according to a PPC Land count of the two texts, which excludes navigation and the changelog. Three headed sections are new: Training and developing models, OpenAI-approved use and Providing access to our models. Most of the introductory text is unchanged, word for word, including the statement that OpenAI reserves "all rights to withhold access where we reasonably believe it necessary to protect our service or users or anyone else," and the privacy section, which still bars uses such as facial recognition databases built without data subject consent, reads exactly as before.

The change drew commentary on the day it took effect. Elena Gurevich, an AI and intellectual property attorney who advises companies on AI governance, wrote in a LinkedIn post on November 20 that "Starting today, Nov 20, OpenAI's updated usage policies kicked in." She summarized the intermediary rule this way: "In simple terms, if you are a business that provides your customers direct access to OpenAI's models, you need OpenAI's approval first. Using these models to power your application's features/functionality doesn't trigger the approval requirement." She gave OpenRouter, which she described as "a platform that streamlines access to multiple AI models from different providers," as an example of the kind of service in scope. That is her reading. OpenAI's policy names no company.

What changed, section by section

Section Version effective October 29, 2025 Version effective November 20, 2026
Licensed advice "provision of tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional" "building or offering AI tools that provide others with tailored advice that requires a license, without appropriate involvement by a licensed professional. You also may not state or imply that the AI itself holds professional credentials or authorization."
State uses needing approval "national security or intelligence purposes without our review and approval" "national security, law enforcement, or intelligence purposes without our review and approval"
Minors "Keep minors safe," one list of 8 prohibited uses "Keep young people safe," split into two subsections with 4 and 5 prohibited uses
High-stakes automation "automation of high-stakes decisions in sensitive areas without human review" "automation of high-stakes actions in sensitive areas without human review," with a definition covering "actions, decisions, transactions, submissions, and other steps"
Sensitive areas listed 13, with "national security" and "law enforcement" as separate entries 12, with "national security, intelligence, and law enforcement" as one entry
Training other models No section New section, including a bar on extracting "protected reasoning"
Case-by-case exceptions No section New "OpenAI-approved use" section
Model access for third parties No section New "Providing access to our models" section requiring "prior review and approval"
Privacy Six prohibited uses Unchanged
Body length (PPC Land count) About 830 words About 1,400 words

The intermediary rule

The provision that prompted Gurevich's post is short. "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," the policy states. "This requirement does not apply solely because an application integrates OpenAI models to power its features or functionality."

Three phrases do the work. "Model-access intermediary platform" names the obvious targets: routers, gateways and aggregators that sit between OpenAI's API and a customer who wants to call a model. "Or other service" widens the net to anything that achieves the same result, whatever it calls itself. And "directly access" is the test that separates a service in scope from the carve-out. An application that uses an OpenAI model to summarize a document or draft an email is integrating the model to power a feature. A service that hands its customers a pipe through which they send their own prompts to an OpenAI model of their choosing looks much more like direct access. Between those two poles sits a large and growing category of products, from agent builders to workflow tools with a model picker, that the text does not place on either side.

The word "prior" also matters. Approval is required before a service of this kind operates, not after the fact. Neither the policy page nor the help notice describes how to apply, how long a review takes, whether services already running on November 20 have a transition window, or whether any intermediary has already been approved.

The rule reverses a direction OpenAI took four years ago. The same changelog records that on October 25, 2022, developers "no longer need to wait for approval after submitting as long as they comply with our policies," and that on November 9, 2022, "We no longer require you to register your applications with OpenAI. Instead, we'll be using a combination of automated and manual methods to monitor for policy violations." The 2026 requirement is narrower than the app review regime OpenAI dropped, since it applies only to services that pass access on to third parties, but for that group it restores a wait for OpenAI's sign-off before operating.

Routers have become a meaningful part of how AI models are consumed. In August, data drawn from OpenRouter showed AI agents using nearly five times as many tokens per task as human users, with agent usage up about 14 times since February 2026 and more than 85% of agent token spend going to cached prompts. A platform that aggregates that volume across providers stands between the model provider and the end customer, and an approval requirement gives OpenAI a view of who those platforms are. OpenAI's documents do not state a rationale for this section.

OpenAI is not alone in tightening control over how access to its models is passed on. When Anthropic's rewritten Usage Policy took effect on November 12, 2026, it added a prohibition on reselling, proxying or otherwise providing access to Claude through unauthorized means, naming services that route requests through consumer subscriptions or misrepresent which product or client is in use.

How the approval provisions are structured

According to the policy text, the new approval mechanisms work in six parts.

  1. Scope. The access rule applies to anyone who operates "a model-access intermediary platform or other service that enables third parties to directly access OpenAI models."
  2. Carve-out. The requirement "does not apply solely because an application integrates OpenAI models to power its features or functionality."
  3. Requirement. Operators in scope "must obtain our prior review and approval." The policy gives no application route, form or review time.
  4. Separate exception route. Under the heading "OpenAI-approved use," the policy states that "OpenAI may grant specific partners case-by-case exceptions to certain restrictions in these Usage Policies." It names the factors reviewed: "the partner, proposed use, applicable safeguards, and the partner's compliance commitments."
  5. Terms of an exception. "These uses may require written terms and apply only to the approved partner, use, and conditions." The policy's worked example is national security, intelligence and law enforcement.
  6. Existing contracts. According to OpenAI's help notice, existing agreements and express permissions in them continue to apply. The training section separately states that its restrictions "do not prohibit uses expressly permitted by your applicable Agreement with OpenAI, including any Permitted Exception under the Services Agreement."

Training rival models

The longest new section opens with a rationale, which is unusual for this document. "Capable models must have safeguards proportionate to the risks they pose before they are deployed," it reads. "Training another model on OpenAI outputs can transfer capabilities without carrying those safeguards over."

The operative prohibition follows: "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." The policy then lists what that covers, in three items:

  • "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 list reads like an inventory of modern post-training practice. Rubrics, preference rankings, rewards and reinforcement-learning environments are the raw materials of the techniques AI labs use to refine models after pre-training, and trajectories are the recorded step-by-step runs of agents. The policy reaches not just the generation of training data but its selection, filtering and validation, which means a model used purely as a judge of someone else's data falls inside the text.

The reach extends beyond the party that trains the model. "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," the policy states. "They also apply when outputs from product use or evaluation are later used or supplied for those training purposes." On a plain reading, that covers data labeling contractors, synthetic data vendors and consultants who never touch the final model, as well as outputs collected during ordinary product use and stored for later.

There are limits in both directions. "You may use OpenAI models for research, software development, model evaluation and benchmarking, and product development, within these boundaries," the section states, and the contract exception covers uses expressly permitted by an agreement with OpenAI. The phrase "Permitted Exception under the Services Agreement" is capitalized as a defined term, but the usage policy does not define it.

Nor does the policy define "alternatives to or substitutes for OpenAI models or services." Read broadly, the phrase could reach a smaller in-house model trained on GPT outputs to take over a task that a company previously sent to OpenAI's API. Read narrowly, it covers competing general-purpose models. The text does not settle which reading OpenAI intends.

Protected reasoning

A single sentence addresses a newer concern. "You must not attempt to extract protected reasoning through unauthorized access, reverse engineering, or circumvention of safeguards," the policy states. The term "protected reasoning" is not defined. Placed inside the training section, it most plausibly refers to the intermediate reasoning that OpenAI's reasoning models produce before an answer and that the company does not show in full, though the policy does not say so.

Distillation in context

Training one model on another's outputs is a documented practice across the industry. PPC Land's coverage of a preprint mapping the ancestry of 15 foundation models recorded that its taxonomy defines model distillation as one model training on another's outputs to mimic it, and that DeepSeek-V3 was distilled from DeepSeek-R1 while Gemma 3's smaller models were produced by distillation. Anthropic has argued publicly for curbs. In July, after 77 organizations signed a letter on open-weights models, Anthropic set out its support for action against industrial-scale distillation, alongside chip export controls and mandatory pre-release testing. Its November 12 policy now bars using outputs to train an AI model without its prior authorization.

The inclusion of "corrections" among the barred training signals echoes an argument made from the other side of the market. In July, Microsoft chief executive Satya Nadella argued that companies buying AI models leak their own know-how through prompts, tool use and corrections, and called on firms to control their own learning loop. OpenAI's rule addresses the flow in the opposite direction: what customers may do with what the model gives them.

Whether such rules hold up in court is a separate question. Stanford's Mark Lemley and Princeton's Peter Henderson argued in a paper dated December 2024 that restrictions AI companies place on their models and outputs may prove unenforceable, because the companies have a weak basis for claiming intellectual property rights in what their systems generate. A usage policy, however, is enforced first through access: OpenAI's text states that breaking or circumventing its rules "may mean you lose access to our systems or experience other penalties."

Children and teens

The child safety section has been renamed from "Keep minors safe" to "Keep young people safe" and split in two. The October 2025 version said OpenAI's services "must never be used to exploit, endanger, or sexualize anyone under 18 years old," and that the company reports "apparent child sexual abuse material and child endangerment to the National Center for Missing and Exploited Children."

The new Child sexual abuse and exploitation subsection states: "Our services must not be used to cause, enable, facilitate, support, or encourage the sexual abuse, exploitation, or sexualisation of anyone under 18." The reporting sentence now refers to "apparent child sexual abuse material and other child sexual abuse conduct," replacing the earlier "child endangerment." The prohibited list keeps child sexual abuse material, "whether or not any portion is AI generated," and expands "grooming of minors" to "grooming, sextortion or any form of conduct based child sexual abuse." Two items are new: "sexualizing children" and "other usage or behaviour that supports the sexual abuse or exploitation of children."

A second subsection, Teens' use of our products and services, states that "OpenAI tools must not be used to endanger anyone under 18" and that the company builds "age-appropriate safeguards designed to reduce exposure to content that may be harmful or developmentally inappropriate." Three items carry over unchanged: exposing minors to age-inappropriate content, promoting unhealthy dieting or exercise behavior, and shaming the body type or appearance of minors. Underage sexual or violent roleplay also remains. Two former items, "dangerous challenges for minors" and "underaged access to age-restricted goods or activities," are merged into one: "promoting or facilitating dangerous challenges or age restricted goods and services to minors." The new wording targets promotion and facilitation rather than access itself, and refers to "goods and services" where the old text said "goods or activities."

The rewrite follows a run of child safety announcements from OpenAI. On November 16, the company said ChatGPT users who try to find child abuse material face 24-hour suspensions and released a free classifier for suspected CSAM, including a signal for images that may be AI-generated, to vetted organizations. On October 30, OpenAI disclosed that it made more than 200 law enforcement referrals in the third quarter of 2026, alongside new limits for teens.

Licensed advice

The licensed advice clause has a history at PPC Land. In November 2025, several outlets reported that ChatGPT would stop giving financial or legal advice after the October 29, 2025 update, and PPC Land traced the reports to a misreading of a clause that had existed before that update. The clause barred "provision of tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional."

The new wording settles the ambiguity that fed those reports by naming who the rule is aimed at. It now bars "building or offering AI tools that provide others with tailored advice that requires a license, without appropriate involvement by a licensed professional." The examples of legal and medical advice have been removed, so the clause no longer names any profession. A second sentence is new: "You also may not state or imply that the AI itself holds professional credentials or authorization." That bars, for example, an AI product that presents itself as a lawyer or a doctor, whatever the involvement of a licensed professional behind it. Regulators had already moved in this area; in January 2025, California's attorney general issued an advisory stating that licensed physicians must supervise AI use that affects medical decisions.

From decisions to actions

The high-stakes rule under Empower people has been rewritten for a world in which AI systems do things rather than just recommend them. The October 2025 version barred "automation of high-stakes decisions in sensitive areas without human review." The new one bars "automation of high-stakes actions in sensitive areas without human review," and adds a definition: "This includes actions, decisions, transactions, submissions, and other steps that materially affect a person's rights, safety, access to essential services, or other consequential interests in areas such as" a list of sensitive fields.

Transactions and submissions are the vocabulary of agents. A week ago, OpenAI gave ChatGPT agents a wallet that completes purchases using network tokens. The rewritten rule makes clear that a human-in-the-loop requirement applies when such steps materially affect a person in a sensitive area, not only when a model renders a verdict.

The list of sensitive areas is otherwise almost unchanged. It covers critical infrastructure, education, housing, employment, financial activities and credit, insurance, legal, medical, essential government services, product safety components, national security, intelligence and law enforcement, and migration. "National security" and "law enforcement," which appeared as separate entries, are merged, and "intelligence" is added, which takes the count from 13 to 12. Three of the listed fields, housing, employment and credit, are the same areas that advertising platforms treat as a special ad category with restricted targeting. The usage policy does not mention advertising, and nothing in it says ad delivery counts as a high-stakes action.

National security and law enforcement

Law enforcement now sits alongside national security and intelligence in the list of purposes that require OpenAI's "review and approval." The new OpenAI-approved use section explains how that approval is meant to work. "Subject to the requirements described above, and as described in our Principles for National Security Partnerships, we support national security, intelligence, and law enforcement uses of our services that benefit people, strengthen democratic governance, and preserve appropriate human judgment and accountability," the policy states.

It then sets limits that apply to every customer. "But all customers and uses must comply with our requirements on preserving human judgment and accountability over high stakes decisions, including the use of force, in the national security and law enforcement space, and with other prohibitions in our Principles on the use of our tools for mass domestic surveillance and evasion of legal obligations, oversight, or accountability." The Principles document itself is referenced but not reproduced on the policy page.

The language lands in a year when the terms on which AI labs work with the US military became a public dispute. In February, Anthropic chief executive Dario Amodei refused Pentagon demands to drop two safeguards, on mass domestic surveillance and fully autonomous weapons, and in March a federal judge blocked the Pentagon from blacklisting Anthropic, finding the government's action likely to be First Amendment retaliation. OpenAI's policy now writes mass domestic surveillance and human judgment over the use of force into its public usage rules by reference to its Principles.

Two policies, eight days apart

OpenAI and Anthropic announced their rewrites 13 days apart and put them into effect eight days apart, and the overlap is substantial.

OpenAI Anthropic
Changes announced October 21, 2026 (changelog) October 8, 2026
Effective date November 20, 2026 November 12, 2026
Passing on model access Prior review and approval for model-access intermediaries; apps that integrate models to power features are exempt Bars reselling, proxying or providing access through unauthorized means, including routing through consumer subscriptions
Training on outputs Bars preparing or supplying training material or signals for alternatives to or substitutes for OpenAI models or services Bars training an AI model on outputs without prior authorization

Anthropic details are from PPC Land's November 12 coverage of that policy.

Why it matters for marketing technology

Marketing software that uses OpenAI models to power its own features, such as drafting copy, summarizing performance data, classifying search terms or answering a customer in a chat window, falls under the carve-out on the text. The rule bites on platforms that sell model access itself, including AI gateways that agencies use to switch between providers, and martech products that let customers bring their own prompts to a choice of underlying models. For those vendors, OpenAI's approval is now a condition of offering its models, and a refusal would remove OpenAI's models from their catalog.

The training section has a wider reach than its framing suggests. Agencies, data vendors and in-house teams that have used GPT outputs to label data, write evaluation rubrics or generate synthetic examples for their own models are now operating under explicit language that covers all of those activities when the resulting model is meant to replace OpenAI's. The policy does not say how OpenAI will establish what a model trained by a third party was "intended" to do.

What the policy does not say

The policy does not define "directly access," "model-access intermediary," "alternatives to or substitutes for," "protected reasoning" or "Permitted Exception." It publishes no application process, review time, list of approved intermediaries or transition period for services already operating, and it does not say how cloud marketplaces that distribute OpenAI models fit in, beyond the help notice's statement that existing agreements continue to apply. That question has grown since OpenAI and Microsoft rewrote their partnership in April, allowing OpenAI to serve its products on any cloud. The Principles for National Security Partnerships are referenced but not reproduced.

Key facts

  • Document: Usage policies, OpenAI
  • Current version: effective November 20, 2026, at https://openai.com/policies/usage-policies/ (also published in UK English at https://openai.com/en-GB/policies/usage-policies/)
  • Previous version: effective October 29, 2025, archived at https://openai.com/policies/usage-policies/revisions/2025-10-29/
  • Changelog entry: dated October 21, 2026; states the changes are effective November 20, 2026
  • Help center notice: "We're updating our Usage Policies" (https://help.openai.com/en/articles/12092907-were-updating-our-usage-policies)
  • New sections: Training and developing models; OpenAI-approved use; Providing access to our models
  • Renamed section: "Keep minors safe" became "Keep young people safe," with subsections "Child sexual abuse and exploitation" and "Teens' use of our products and services"
  • Sensitive areas for high-stakes automation: 13 in the October 2025 version, 12 in the new version
  • Defined terms referenced but not defined in the policy: Agreement, Permitted Exception, Services Agreement
  • Referenced document: Principles for National Security Partnerships
  • Changelog history: 14 entries, from February 26, 2021 to October 21, 2026
  • Commentary: LinkedIn post by Elena Gurevich, AI and IP attorney, November 20, 2026

Timeline

Summary

Who: OpenAI, and every person or business that uses its services, with specific obligations for operators of platforms that give third parties direct access to OpenAI models, for anyone using OpenAI outputs to prepare training material for other models, and for national security, intelligence and law enforcement users. Elena Gurevich, an AI and IP attorney, commented on the change.

What: A rewritten Usage Policies document that requires prior review and approval for model-access intermediaries, bars using OpenAI models or outputs to generate, select or supply training material or signals for rival models, bars attempts to extract protected reasoning, adds a case-by-case exception route, rewrites the child safety section, extends the high-stakes rule from decisions to actions, adds law enforcement to the uses needing approval, and narrows the licensed advice clause to those building or offering AI tools.

When: The changes were announced in a changelog entry dated October 21, 2026 and took effect today, November 20, 2026, replacing the version effective October 29, 2025.

Where: The Usage Policies apply across all OpenAI products and services, a universal scope OpenAI adopted in its October 29, 2025 update.

Why: According to OpenAI's changelog, the update clarifies existing restrictions on training other models and existing protections, and incorporates its national security principles. The policy's own rationale for the training section is that training another model on OpenAI outputs "can transfer capabilities without carrying those safeguards over." OpenAI gives no stated reason for the intermediary rule.