OpenAI Will Require Model Routers and AI Gateways to Get Its Approval Before Offering Direct Access to Its Models, From November 20
OpenAI's revised usage policies, effective November 20, say any service that lets third parties directly access OpenAI models needs its prior review and approval, a rule OpenAI says covers model routers and AI gateways. The update also bars using OpenAI outputs to train substitute models and adds law-enforcement uses to those needing approval. Neither the policy nor OpenAI's post says how approval works.
OpenAI is giving itself a say over which third-party services can offer direct access to its models. Its revised usage policies, published on Wednesday, October 21 and effective November 20, 2026, add a section headed "Providing access to our models": "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 adds that the requirement "does not apply solely because an application integrates OpenAI models to power its features or functionality."
What OpenAI says the rule is for
In a post titled "Broad access through consistent standards," OpenAI says the update means "requiring services that provide direct access to our models such as model routers or AI gateways to get our review and approval before offering that access." It says partners that provide access must "maintain safeguards consistent with our own, investigate potential violations, and take appropriate action when concerns are identified," including coordinating with OpenAI on enforcement. OpenAI also says it is working with customers on SafetyIDs, which help it "distinguish between individual end users in an investigation without OpenAI learning those users' identities."
Training on OpenAI outputs
The revised policies say users "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." That covers creating training data, tasks, rubrics or reinforcement-learning environments; giving feedback such as labels, scores, rankings, preferences, rewards or corrections; and selecting, filtering or validating training data. Users also "must not attempt to extract protected reasoning," and the restrictions apply "even if you do not train the resulting model." Research, software development, model evaluation and benchmarking remain allowed within those limits, as do uses expressly permitted by a customer's agreement with OpenAI.
The list of prohibited uses now includes "national security, law enforcement, or intelligence purposes without our review and approval." OpenAI's changelog says the update "sets out approval requirements for national security and law enforcement uses, as well as for model-access intermediaries." Mad Robot, a tech news site, says the previous version required review for national security or intelligence work.
What the policy leaves open
Mad Robot names OpenRouter and Vercel's AI Gateway as examples of services that fit OpenAI's description, adding that "OpenAI doesn't name any of them, and none has commented publicly on the change." It says 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," nor whether open-source gateway software that companies run for their own staff counts. Mad Robot gives LiteLLM as an example of that software and points to its compromised release on October 21 as a sign that gateways are a security weak point. It also links the timing to OpenAI and Google's announcement the same day that its models are coming to Google Cloud's Model Garden.
Limits
- The rule takes effect on November 20. The policy and OpenAI's post we read do not describe an application process, a review timeline, or what happens to services already offering access.
- The examples of affected services are Mad Robot's; OpenAI names none. The only outside coverage we read was Mad Robot's.
- We did not compare the new text with OpenAI's previous version; the law-enforcement change is described by OpenAI's changelog and Mad Robot.
Sources
- OpenAI: Usage policies (effective November 20, 2026; changelog entry October 21, 2026)
- OpenAI: Broad access through consistent standards (October 21, 2026)
- Mad Robot: AI routers like OpenRouter will soon need OpenAI's permission to sell GPT models (October 22, 2026)
- Times of AI: LiteLLM's Docker and GHCR images were also poisoned (October 21, 2026)
- Times of AI: OpenAI models and Codex are coming to Google Cloud's Model Garden (October 21, 2026)