OpenAI safety policies are changing. On 21 October 2026, OpenAI announced an update to the rules that govern how its AI models may be used — and who may pass access to those models on to others. The new Usage Policies take effect on 20 November 2026, and their headline change is a requirement that services offering direct access to OpenAI models, such as model routers and AI gateways, get OpenAI’s review and approval first.
The update is the written form of work OpenAI says it has been doing for the past few months: aligning safety policies across every service that provides its models. GPT-6 Astra is named in the announcement as an example of the models the aligned policies cover. For developers and businesses, the practical question is what actually changes, and what they need to do before 20 November.
What OpenAI announced
OpenAI announced the changes in a blog post, “Broad access through consistent standards”, published on 21 October 2026, alongside the revised Usage Policies themselves. The announcement has two parts. First, OpenAI says it has spent the past few months working with customers and partners to align safety policies across the services that provide its models, so that the same safeguards apply wherever a developer or business accesses them. Second, it has updated its Usage Policies to put that alignment in writing and to clarify the rules for safe model development and sensitive uses of AI.
According to OpenAI’s Help Center, the refresh contains four updates: approval requirements for model-access intermediaries; clarified restrictions on using OpenAI models and outputs to train models meant as alternatives to OpenAI’s; clarified protections for children and teens and requirements for high-stakes uses; and national security principles with approval requirements for national security and law enforcement uses. The changes take effect on 20 November 2026 — thirty days after the announcement — which gives partners roughly a month to adjust. One caveat is worth noting: for paying business customers, the business contract decides which policy version applies, so 20 November is not a universal deadline for every customer.
One set of safety standards, everywhere you access the models
The core idea of the announcement is that the rules should follow the model, not the channel. Developers and businesses should be able to choose where they access OpenAI’s models, with consistent safeguards across those services — whether they go directly through OpenAI, through a partner that hosts the models, or through a reseller. OpenAI says it developed these policies by working with customers and partners to align safeguards across providers.
That alignment comes with obligations for the partners. Those that provide access to OpenAI models must maintain safeguards consistent with OpenAI’s own, investigate potential violations, and take appropriate action when concerns are identified — including coordinating on enforcement, so that a restriction imposed in one place cannot simply be dodged by switching to another service. The announcement points to OpenAI’s work with customers and partners on Private Safety Processing, which offers zero data retention for frontier models, as an example of how safeguards can be developed while giving customers greater choice and control over their sensitive data.
The safeguards themselves start before a model is deployed. OpenAI says it evaluates a model’s capabilities and risks before deployment, combines safeguards designed to reinforce one another, and tests them. After deployment, it monitors for potential abuse and investigates concerns. Trusted access programmes provide another layer of protection, setting eligibility requirements and conditions for access to its most powerful capabilities. OpenAI models are also widening their distribution: access through Google Cloud’s Model Garden and Marketplace is planned, with general availability in the coming weeks.
The updated Usage Policies: what is changing
Beyond the intermediary rule, the revised Usage Policies clarify several areas that matter to anyone building on OpenAI’s models.
The most detailed change concerns training other models. OpenAI now states plainly that you must not use its 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. That covers creating or preparing training data, training tasks, rubrics, or reinforcement-learning environments, and providing feedback used in training — including labels, scores, rankings, preferences, rewards, or corrections. The restriction applies 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 reason OpenAI gives is that training another model on its outputs can transfer capabilities without carrying the safeguards over. Research, software development, model evaluation, and benchmarking remain allowed within those boundaries. The policies also bar attempting to extract protected reasoning through unauthorised access, reverse engineering, or circumvention of safeguards.
Protections for children and teens are clarified. The policies prohibit using OpenAI’s services to expose minors to age-inappropriate content, to promote unhealthy dieting or exercise behaviour to minors, and for underaged sexual or violent roleplay.
Requirements for high-stakes uses are clarified too. The policies do not allow automation of high-stakes actions in sensitive areas without human review, and the updated text explicitly includes consequential actions, transactions, and submissions. The areas listed include critical infrastructure, education, housing, employment, financial activities, insurance, legal, medical, essential government services, national security, and migration.
Finally, national security, law enforcement, or intelligence purposes now require OpenAI’s review and approval, with law enforcement newly added alongside national security and intelligence. Anthropic, for context, announced its own usage policy update on 8 October 2026, effective 12 November, covering deceptive campaigns, oversight of high-risk uses, and autonomous hardware — a sign that the major labs are tightening these rules in parallel.
OpenAI safety policies: model routers and AI gateways now need approval
The headline change is the new rule for intermediaries. The Usage Policies now state:
“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. This requirement does not apply solely because an application integrates OpenAI models to power its features or functionality.”
Model routers and AI gateways are examples of the intermediaries the rule covers: services that provide third parties with direct access to OpenAI’s models. Under the new rule, any service that sits in the middle and passes direct access to OpenAI’s models along to third parties needs OpenAI’s sign-off before offering that access.
The test is whether a service enables third parties to access the models directly. The policy does not apply solely because an application integrates OpenAI models to power its own features — a business that uses them to summarise documents or draft support replies inside its own product, for example — but an application that also gives third parties direct access to the models is not exempt on that basis alone. That is what decides who has to act before 20 November, subject to the caveat that a business customer’s contract decides which policy version applies to it.
What OpenAI has not said is how the approval process will work. The announcement names no router or gateway company, and the published policy text sets out no application route, criteria, or timeline — leaving open what standards operators will be judged against, how long approval takes, and what happens to existing unapproved intermediaries on 20 November. Until the process is published, operators cannot plan around it with confidence.
SafetyIDs: tracing misuse without identifying users
Alongside the policy changes, OpenAI says it is working with customers to introduce SafetyIDs. These are identifiers that customers can provide to help OpenAI distinguish between individual end users in an investigation, without OpenAI learning those users’ identities. The point is precision: when misuse is traced to one end user, OpenAI can respond to that misuse specifically, rather than restricting an entire customer account and disrupting everyone else who uses it.
The need for that precision follows from how misuse reaches OpenAI’s customers: through their end users, compromised accounts, or leaked API keys. SafetyIDs are designed for the first of those vectors — the end user whose behaviour triggers an investigation. Whether they are opt-in, and how customers implement them technically, is not yet clear from the announcement.
How OpenAI enforces the rules
OpenAI describes its enforcement approach in the announcement as follows:
“We base enforcement decisions on documented conduct and risk, apply our rules consistently, and respond in proportion to the circumstances.”
OpenAI also says it revisits a restriction when meaningful changes in conduct, commitments, or safeguards address the concerns behind the decision. The Usage Policies add that users can appeal if they think OpenAI has made a mistake enforcing policy. Breaking or circumventing the rules and safeguards can mean losing access to the systems.
Enforcement, in OpenAI’s framing, is part of the same consistent-standards logic as the rest of the update: partners must investigate potential violations and coordinate with OpenAI, so that a restriction imposed on one service cannot be evaded by moving to another. The aim, as the announcement closes, is that clear rules and consistent safeguards should make it easier for developers and businesses to build, compete, and choose the services that work best for them.
What developers and businesses should do now
The update leaves different readers with different homework. If you access OpenAI’s models through a model router or an AI gateway, check with the provider whether it has OpenAI’s approval — or is seeking it — before 20 November. If you operate such a service, the approval requirement applies to you directly, and the absence of a published process means you should be asking OpenAI how to apply sooner rather than later.
If you build applications on OpenAI’s models, read the updated Usage Policies before they take effect, and check your contract for which policy version applies to you. Pay particular attention to the training rules if you build or fine-tune models of your own: using OpenAI’s outputs to train a model intended as an alternative is now explicitly barred, while research, evaluation, and benchmarking remain allowed. Review your API key handling too — leaked keys are one of the misuse vectors OpenAI names.
If you serve end users through your own product, watch for SafetyIDs. Once OpenAI publishes how they work, providing them could mean that a misuse investigation affecting one of your users stays targeted at that user, instead of disrupting your whole account. And if your product touches children, teens, or high-stakes decisions — hiring, credit, housing, medical advice, or legal work — check the clarified protections and the human-review requirements against what your product actually does.
The through-line of the whole update is simple: the safety rules now follow the model wherever it goes, and any service that gives third parties direct access to OpenAI’s models must first get OpenAI’s review and approval — while an application that merely integrates the models into its own features is not covered by that requirement. What remains to be seen is how strictly OpenAI reviews those middlemen — and whether the approval process, once published, feels like a fair standard or a new gatekeeper.
