OpenAI will require outside platforms that resell or route direct access to its models to obtain its approval before Nov. 20, under a revised usage policy the company announced on Oct. 21. The change targets so-called AI gateways and model routers — services that let a business plug into several different AI systems through one connection instead of integrating each model separately — and marks one of OpenAI's most concrete steps yet to control how its technology spreads through intermediaries it does not directly operate.
The approval requirement draws a line between two kinds of products, according to OpenAI's updated terms described in OpenAI's help center and detailed further by eWeek's review of the policy. An application that simply builds GPT models into its own product, such as a customer-support assistant or a document-generation tool, is not covered. A service whose business is routing or reselling access to the underlying models — letting a customer swap between OpenAI and a rival provider through one interface, for instance — now needs sign-off.
Safety rationale, few specifics
OpenAI has framed the requirement as a safety measure rather than a competitive one. The company says it wants consistent safeguards applied no matter which intermediary a customer uses, so that switching providers cannot become a way to dodge restrictions placed on direct OpenAI access. Approved partners will need to show safeguards comparable to OpenAI's own, maintain a process for investigating suspected violations and coordinate with OpenAI when enforcement action is needed. OpenAI is also developing a system of "SafetyIDs" intended to help distinguish individual end users during an investigation without revealing their identity to OpenAI itself, though the current policy does not require every gateway to adopt it by the Nov. 20 deadline.
What OpenAI has not published is an approved-gateway list or a clear process for how an existing intermediary applies for review before the deadline arrives, leaving gateway operators to seek guidance directly from OpenAI about whether the rule applies to them. The policy revision also tightened separate restrictions on model distillation — the practice of using OpenAI's outputs to train a rival model — and added approval requirements for AI use in security and law-enforcement contexts, changes OpenAI bundled into the same update.
The practical effect falls on a layer of the AI industry most consumers never see: the gateway and router startups that have proliferated as businesses try to hedge against being locked into a single model provider. Any of those services found to be operating without approval after Nov. 20 could see their access to OpenAI's models disrupted, giving them a narrow window to either seek sign-off or restructure how they offer multi-model access to customers.
