OpenAI has updated its Usage Policies so that any service that lets third parties directly access OpenAI models must get OpenAI's prior review and approval. The change takes effect on November 20, 2026. It matters most to developers who reach OpenAI models through a router or gateway, and to the operators of those services. An app that simply builds OpenAI models into its own features does not trigger the requirement on that basis alone.

Who needs approval

The new clause, titled "Providing access to our models," 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."

In its blog post, OpenAI gives model routers and AI gateways as examples of services that provide direct access to its models. The clause sits in the policy's "OpenAI-approved use" section, where OpenAI says it may grant specific partners case-by-case exceptions after reviewing the partner, the proposed use, the safeguards and the partner's compliance commitments.

Who is not affected

The next sentence sets the boundary: "This requirement does not apply solely because an application integrates OpenAI models to power its features or functionality."

The word "solely" matters. Integrating OpenAI models into a product does not by itself create a need for approval. It is not a blanket exemption, though: an app that also lets third parties directly access OpenAI models falls under the first sentence. Under the policy's wording, a developer that calls the OpenAI API to power its own product doesn't need approval for that reason alone.

When it starts

The policy page lists November 20, 2026 as its effective date, and the changelog entry is dated October 21, 2026. OpenAI has not said how to apply for approval, what criteria it will use, how long reviews take, or what happens to services already operating on November 20.

Training restriction and SafetyIDs

The update also makes the training rules explicit. 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." The policy lists creating training data, tasks, rubrics or reinforcement-learning environments; providing feedback such as labels, scores, rankings, preferences or rewards; and selecting, filtering or validating training data. It also bans attempts to extract protected reasoning through unauthorized access, reverse engineering or circumventing safeguards.

These limits apply "whether you perform the work yourself, for another party, or through someone acting on your behalf, even if you do not train the resulting model." Research, software development, evaluation, benchmarking and product development remain allowed within these limits, as do uses expressly permitted by a customer's agreement with OpenAI.

The blog post adds that partners providing access to OpenAI models must keep safeguards consistent with OpenAI's own. OpenAI is also introducing SafetyIDs with customers, which customers can provide so OpenAI can tell individual end users apart in an investigation without learning who they are.

The same update also clarifies protections for children and teens and rules for high-stakes uses, and requires OpenAI's approval for national security and law enforcement uses.

For more on how OpenAI prices access to its models, see the GPT-6.1 Sol Ultrafast pricing breakdown.