The OpenAI logo. Image: OpenAI (logo, public domain via Wikimedia Commons)
OpenAI is about to start deciding who gets to resell its AI. In updated Usage Policies that take effect on November 20, any service that lets other people “directly access OpenAI models” will first need the company’s “prior review and approval”. In a blog post on Wednesday, OpenAI said the rule covers “model routers or AI gateways”, the services developers use to reach GPT-6 Astra and hundreds of rival models through a single account.
That describes services such as OpenRouter and Vercel’s AI Gateway, which sell access to OpenAI’s models alongside those from Anthropic, Google, Meta and others. OpenAI doesn’t name any of them, and none has commented publicly on the change.
What the new rule says #
The clause sits in a new section called “Providing access to our models”. It applies to anyone who operates “a model-access intermediary platform or other service that enables third parties to directly access OpenAI models”. Ordinary apps are left alone: the rule “does not apply solely because an application integrates OpenAI models to power its features or functionality”. So a writing app that uses GPT-6 behind the scenes doesn’t need sign-off, but a service that hands its customers raw access to OpenAI’s models does.
The policy doesn’t say how a router applies for approval, how long a review takes, or what happens to services already offering OpenAI models if they haven’t been approved by November 20. It also doesn’t say whether open-source gateway software that companies run for their own staff, such as LiteLLM, counts.
Same safeguards, wherever you buy #
OpenAI frames the change as part of wider work with partners. Companies that host or resell its models must “maintain safeguards consistent with our own, investigate potential violations, and take appropriate action”, it says, including “coordinating on enforcement so that restrictions remain effective across services”. In other words, a developer banned on OpenAI’s own platform shouldn’t be able to carry on through a reseller.
The timing is pointed. Earlier the same day, OpenAI and Google said GPT-6 Astra is coming to Google Cloud, putting OpenAI’s models on all three of the biggest cloud platforms. And outside researchers showed last month that a “reasoning replay” attack OpenAI had already blocked on its own API still worked on Microsoft Azure for days afterwards, concluding that labs can’t treat resellers as “passive distribution channels”.
Gateways are also a security weak point. On Wednesday, a hijacked release of LiteLLM, a popular gateway tool, shipped a credential stealer that raided the provider keys stored inside it. OpenAI says customers face misuse “through their end users, compromised accounts, or leaked API keys”, and it is introducing SafetyIDs, which let a customer label individual end users so OpenAI can act against one bad actor “without OpenAI learning those users’ identities”.
Copying and policing get new rules too #
The router clause isn’t the only change. OpenAI’s terms of use already bar using its output “to develop models that compete with OpenAI”; the new Usage Policies spell that out in detail. Customers must not use OpenAI models to create training data, reinforcement-learning environments, rubrics or feedback signals for rival models, and the ban applies “even if you do not train the resulting model”. Trying to extract the models’ hidden reasoning is also banned. That follows OpenAI’s accusation that people linked to Kimi maker Moonshot AI tried to copy its reasoning, and Anthropic’s claim that its own defences now cut what copycats harvest from Claude by up to 99%.
The update also adds law enforcement to the uses that need OpenAI’s approval. The previous version required review for national security or intelligence work; the new one covers “national security, law enforcement, or intelligence purposes”, with exceptions granted case by case under OpenAI’s national security principles.
Why it matters #
Routers are how many developers try out OpenAI’s models next to its rivals’, so OpenAI now holds a veto over one of the main roads to its own customers. It says the aim is consistent safety, but the same power lets it pick which middlemen survive, and so far it hasn’t published the criteria.
Sources: OpenAI, “Broad access through consistent standards”; OpenAI Usage Policies (effective November 20, 2026); OpenAI Usage Policies, previous version; OpenAI Terms of Use.