OpenAI's updated usage policies quietly became binding on November 20. The changes were published in the policy's changelog on October 21, which states the update sets out "approval requirements for national security and law enforcement uses, as well as for model-access intermediaries," and that the changes took effect on November 20. The practical consequence for the developer ecosystem is large, even though the rollout arrived without a launch blog post.
The first big change targets model gateways. The policy now reads: "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 requirement does not apply to ordinary applications that integrate OpenAI models to power their own features, so a CRM that uses GPT for drafting emails is exempt, while a router that resells raw API access is not.
That distinction lands directly on a fast-growing layer of the ecosystem. OpenRouter, the largest model aggregator, was acquired by Stripe in a reported 7.5 billion dollar deal that completed earlier this year, and Hugging Face has since handed its model-router traffic to the same platform. Under the new policy, gateways of this kind become businesses that operate on OpenAI's approval rather than on generic API terms.
The second change closes the distillation loophole in writing. The policy now states: "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." It adds that attempting "to extract protected reasoning through unauthorized access, reverse engineering, or circumvention of safeguards" is prohibited, and that the restrictions apply even when the work is done through a contractor, "even if you do not train the resulting model."
There are explicit carve-outs. OpenAI still permits using its models for research, software development, model evaluation and benchmarking, and product development, and says the restrictions do not override anything expressly permitted by a customer's Services Agreement, including defined Permitted Exceptions. The gap between those two categories, allowed evaluation versus banned data generation, is where most compliance arguments will happen.
The rules mirror a direction Anthropic has already taken, with media reports in November describing Claude policy changes that restrict unauthorized reselling and training on outputs. The two largest Western labs are converging on the same stance: model access is a licensed, auditable relationship, not a commodity that can be arbitraged.
What OpenAI has not published is how approval will work in practice. There is no public application process, no stated criteria, and no answer to whether gateways must disclose their end customers. Aggregators that route traffic across many labs now face a choice: seek approval and accept audit, migrate customers to direct enterprise agreements, or lean harder on open-weight models that carry no such conditions. The last option is probably the policy's real side effect, and the one OpenAI's competitors will be happiest about.
Comments (0)
Log in to join the discussion
Log InNo comments yet