OpenAI tightens Usage Policies for model routers and resale partners

On October 21, OpenAI published a post titled "Broad access through consistent standards" describing months of work with customers and partners to align safety policies across the services that offer its models, and announcing an update to its Usage Policies. The revised policies take effect on November 20, 2026, and are meant to make the rules clearer and easier for developers to follow.

The most concrete change targets the reseller layer. Services that provide direct access to OpenAI models — model routers and AI gateways among them — must now obtain OpenAI's review and approval before offering that access.

Partners also have to keep safeguards consistent with OpenAI's own, investigate potential violations, take appropriate action when concerns surface, and coordinate on enforcement so that restrictions hold across every service that provides access to the models.

OpenAI frames the requirements as a consequence of capability growth. Before deployment it evaluates a model's capabilities and risks, trains it to follow safety requirements, and tests whether safeguards survive deliberate attempts to bypass them; after launch it monitors for abuse patterns and investigates specific concerns. Trusted access programs add eligibility conditions for the most powerful capabilities.

The post is explicit that not all protection lives inside the model. Some safeguards are built in, while others depend on the systems and people around it, and OpenAI argues both layers must work whether a user reaches a model directly or through a partner.

As an example of pairing safeguards with customer choice, OpenAI points to Private Safety Processing work with customers and partners, which it describes as a way to keep protections effective while giving customers more control over sensitive data.

The bigger signal is that distribution is being treated as a compliance surface in its own right. Gateway and routing vendors now have a review step on their roadmap, and enterprise buyers inherit a longer vendor chain to verify: they need to know whether the channel they use has been approved.

What to watch next: how OpenAI handles and documents approvals before the November 20 effective date, and whether multi-model gateways can meet a consistent-safeguards bar without giving up the flexibility that made them popular.