For as long as developers have had a choice of models, a layer of middleware has grown up to manage that choice: routers that send each request to the cheapest or fastest provider, gateways that hold keys and enforce spending limits, resellers that bundle access into a bigger product. From Nov. 20, that layer needs OpenAI’s permission.
The company said Wednesday that its updated usage policies, effective Nov. 20, require services that provide direct access to its models, “such as model routers or AI gateways,” to obtain OpenAI’s review and approval before offering that access. The rules apply to customers using OpenAI directly, to partners that host its models, and to reseller partners.
The policy also obliges partners to maintain safeguards consistent with OpenAI’s own standards, to investigate potential violations and to coordinate enforcement with the company. It extends an existing prohibition on using OpenAI model outputs to train competing systems, and it introduces a technical mechanism called SafetyIDs that lets OpenAI separate end users during an investigation without learning who they are.
That last detail is the most concrete thing in the announcement and the one with the clearest operational cost. A gateway that routes traffic for thousands of corporate customers currently presents OpenAI with a single account and a blur of requests. SafetyIDs give the company a way to attribute behavior to a distinct end user, in the way a bank attributes transactions to an account number rather than a name, which is precisely what an abuse investigation needs and precisely what a reseller’s customers may not want disclosed.
The middleware layer exists because enterprises do not want to rewrite their applications every time a model changes. A single integration point lets a company compare vendors, fail over during an outage, cap spending per team and keep prompts and data flows in one place. Some of these gateways are commercial products with their own pricing; others are open-source projects that a platform team runs internally.
Routing is no longer a niche. It became the default architecture for companies that wanted several models in production at once, cheap ones for bulk classification and expensive ones for reasoning, without rebuilding the product each time a vendor shipped an upgrade. That arrangement depends on the gateway staying neutral between suppliers, which is exactly what an approval regime strains.
OpenAI’s post did not name which gateways are affected, did not define what counts as “direct access,” and did not say how long approval takes. Those gaps matter more than they might appear. A router that passes requests through to OpenAI’s API and one that pools capacity across a dozen providers are different businesses, and the policy language does not draw the line between them.
Enforcement is described in procedural terms. OpenAI said it will base decisions on documented conduct and risk, apply rules consistently, respond proportionately, and revisit a restriction if changes in conduct or safeguards address the concern. There is no appeals timeline, and the commitment reads as policy rather than contract.
Analysts said the immediate burden falls on smaller operators, who will need compliance staff and legal review for a business whose margins are thin. “The large platforms can absorb a review cycle,” said one analyst who follows developer tools. “A five-person gateway startup cannot afford to be offline for a month waiting for an answer.” Larger cloud providers and enterprise software vendors already negotiate bespoke terms and are less exposed.
The move also fits a pattern across the industry. Model providers have spent two years tightening the perimeter around their APIs as agents began acting on behalf of users, executing code and moving money. A router is an attractive place to hide abuse, because it can spread traffic across accounts and obscure who is behind a request. Requiring approval turns an anonymous pipe into a known counterparty.
For buyers, the practical question is contractual. Companies that route production traffic through a third-party gateway should expect new terms, new disclosure requirements and possibly new latency as identity checks are added. Procurement teams that assumed model access was a commodity purchased through any intermediary now have to map which intermediaries are approved, and what happens to their application if one is not.
There is a competitive angle too. Approval gives OpenAI a veto over how its models are packaged, which is useful when a reseller’s pricing undercuts the company’s own list price or when a bundle hides which model is answering. It also gives OpenAI visibility into demand it currently sees only in aggregate.
OpenAI has not said whether approval will be granted on standard terms or negotiated case by case, and it has not said what happens to a gateway that applied before the deadline and is still waiting when the rules take effect.
If a router must be approved to exist, who ends up owning the switch that decides which model answers a question?
