Routers and AI gateways will need OpenAI’s sign-off

Starting November 20, 2026, model routers, AI gateways and other services that let third parties call OpenAI models directly will need OpenAI’s review and approval before they offer that access. OpenAI announced the rule on October 21 as part of updated Usage Policies that take effect on that date.

The approval rule is the most concrete part of a broader push, described in OpenAI’s post “Broad access through consistent standards”, to apply the same safeguards wherever its models are served. OpenAI says it has spent the past few months working with customers and partners to align safety policies across those services. The policies cover direct access through OpenAI, partners that host the models and reseller partners. OpenAI says models such as GPT-6 Astra expand what developers and businesses can build, and it wants safeguards to stay effective as those capabilities grow.

A router or gateway sits between an application and one or more model providers. The developer calls a single endpoint, and the service decides which model gets the request and handles keys and billing. From November 20, the service that picks the model will itself need OpenAI to pick it.

Hosting and reseller partners get their own obligations. They have to maintain safeguards consistent with OpenAI’s, investigate potential violations and act when they find a problem. OpenAI says it is working with each partner to make sure their implementation “meets or exceeds our safety standards.” Partners also have to coordinate with OpenAI on enforcement, so a restriction stays in force across every service that provides access to its models. A customer cut off on one route should not be able to switch to another and carry on.

OpenAI describes two layers of safeguards. Some are built into the model: it is trained to follow safety requirements and refuse harmful requests, and before release it is tested to see whether its safeguards hold when people deliberately try to get around them. Others depend on the systems and people around the model. For cyber misuse, for example, OpenAI runs systems that look for suspicious patterns across interactions to decide how to respond. Trusted access programs set eligibility requirements and conditions for the most powerful capabilities. OpenAI says both layers need to work whether people reach its models directly or through a partner.

The updated Usage Policies also clarify the rules for safe model development and for sensitive uses of AI. OpenAI says it will keep explaining what is allowed, what is restricted, and how customers and partners can raise questions or resolve compliance concerns.

Two parts of the announcement are aimed at customers. The first is Private Safety Processing, which OpenAI presents as an example of keeping safeguards effective while giving customers more choice and control over their sensitive data. The second is SafetyIDs, which OpenAI says it is still working with customers to introduce. A customer passes an ID for each end user, so in an investigation OpenAI can tell those users apart without learning who they are. OpenAI says this lets it respond to potential misuse “more precisely while reducing disruption to others.”

OpenAI also names where misuse tends to come from. Customers can run into it through their own end users, compromised accounts or leaked API keys, and OpenAI says it works with them to understand and fix those problems. On enforcement, it says decisions rest on “documented conduct and risk,” that rules are applied consistently, and that responses are proportionate to the circumstances.

Restrictions do not have to be permanent. When OpenAI restricts access, it says it will revisit the decision “if meaningful changes in conduct, commitments, or safeguards address the concerns behind our decision.”