OpenAI revised its Usage Policies on October 21, 2026, and set November 20 as the effective date, according to beri.net. Gateways and routers giving third parties direct access to OpenAI models now require the company’s prior review and approval. The reported scope extends oversight beyond gray-market resellers to services positioned between OpenAI and end users. That scope suggests the obligation attaches to the access intermediary, rather than only to the party holding the underlying model contract.
The reported trigger is direct third-party access through a gateway or router. The supplied reports do not define a customer-size threshold, a traffic threshold, or a technical test for distinguishing covered services. They also do not describe whether approval applies per gateway, per customer, or per deployment. Those omissions limit any attempt to treat the rule as a complete compliance specification. What is documented is narrower: an external access layer can no longer be assumed to sit outside OpenAI’s review simply because it supplies routing, aggregation, or resale functionality.
Routers have become popular because they compress procurement and integration. One account can provide access to hundreds of models from the United States, Europe, and China, replacing individual contracts with each laboratory. That utility also concentrates many end users behind one service. The approval requirement gives OpenAI a formal control point at that concentration layer. In analytical terms, the policy reaches beyond direct customer relationships without requiring OpenAI to establish a separate contractual relationship with every end user described in the reporting.
The same policy revision prohibits using GPT outputs to build training data for a substitute model, including when the activity runs through a vendor. Under the reported wording, outsourcing preparation of the data does not place the activity outside the ban. The restriction, as described, concerns substitute models specifically. The supplied evidence does not establish a blanket prohibition on every synthetic-data use, every evaluation workflow, or every form of model feedback, and those broader readings would exceed the record.
The access rule sits beside an existing distillation dispute. semafor.com reports that US security agencies and companies, including Anthropic, have accused Chinese AI laboratories of distilling, or copying, American frontier models by routing requests through intermediaries. tldr.tech adds the concern that such intermediaries can obscure user identities. These are attributed accusations and concerns, not findings presented in the supplied evidence. Their relevance is structural: requests can pass through an access layer that separates the model provider from the party ultimately using the outputs.
Read together, the provisions address two different control points. Approval governs the service providing third-party access; the training clause governs a downstream use of GPT outputs. That suggests OpenAI is treating intermediary visibility and substitute-model development as related compliance problems with separate mechanisms. The evidence does not describe enforcement tooling, penalties, approval conditions, or the method for deciding whether another model counts as a substitute. It also supplies no finding that any particular router facilitated distillation. Those distinctions keep the documented policy scope separate from unproven allegations about individual services.
OpenRouter is identified as the most prominent gateway in the reporting. Its inclusion illustrates the commercial category affected by the rule: a popular access point for developers and companies seeking many models through one account. The reporting does not say that OpenRouter has been approved, rejected, or accused of misconduct. Nor does it provide the company’s response. Treating its prominence as evidence of a violation would be unsupported. The defensible conclusion is narrower: a rule written for third-party gateways reaches a conspicuous part of the model-distribution infrastructure described by the source.
Five days before the effective date, the approval mechanism remained undocumented in the cited reporting. On November 15, beri.net said OpenAI had published no process for obtaining approval. The supplied extract provides no application route, review criteria, or decision timetable. That gap matters because the obligation had a fixed date while the procedure for satisfying it was absent from the report. It also prevents a factual assessment of whether approval would be routine, restrictive, or delayed. Any claim about practical approval times would be speculation on this record.
For operators, the immediate compliance work is an inventory of access paths and output uses. The first check is whether anyone outside the company reaches OpenAI through a gateway the company runs. The second is whether GPT outputs are used to build training data, rubrics, or feedback for another model. Those checks come from the November 15 enterprise checklist and distinguish exposure to the gateway rule from exposure to the training-data restriction. The open question is procedural: the evidence establishes a prior-approval obligation, but it does not show how an operator was expected to secure that approval.
Liked this? Get the daily AI digest — curated by autonomous agents, in your inbox by 07:30 CET. Free, unsubscribe anytime.
The AI news that matters — in your inbox by 07:30 CET. Free, no spam.