OpenAI to require approval for model gateways from November 20, 2026

OpenAI frames the policy update as an extension of work already underway: over the past few months it has been aligning safety policies with customers and partners across services that provide its models. The company describes a layered approach — evaluating a model's capabilities and risks before deployment, training it to follow safety requirements, testing whether safeguards hold when people try to bypass them, and monitoring for abuse afterward. Its point is that some safeguards sit inside the model while others depend on the systems and people around it, so both need to work whether access comes directly or through a partner.

The new requirements extend that logic to intermediaries. Partners providing access to OpenAI models must maintain safeguards consistent with OpenAI's own, investigate potential violations, and take appropriate action, including coordinating on enforcement so restrictions remain effective across services. The Usage Policies add a procedural gate for services offering direct access, such as model routers or AI gateways, which must obtain OpenAI's review and approval before doing so.

OpenAI also points to practical mechanisms for handling misuse, which customers can encounter through end users, compromised accounts, or leaked API keys. SafetyIDs are intended to let OpenAI distinguish between individual end users in an investigation without learning their identities, which the company says allows more precise responses with less disruption to others. It says enforcement decisions rest on documented conduct and risk, and that restrictions are revisited if meaningful changes in conduct, commitments, or safeguards address the original concerns.