The News, Explained
On October 21, 2026, OpenAI announced safety standards across model-access services and revised Usage Policies. The revision takes effect November 20 and sets out prior review and approval for services providing direct model access. The announcement names model routers and AI gateways: intermediary services that pass users’ requests to models and provide access to them. Source
OpenAI describes safeguards inside models alongside protections implemented by the services around them. Its approach covers direct access, partners hosting models and reseller partners. It says partners must maintain safeguards, investigate potential violations and coordinate appropriate enforcement. The company is working with partners on implementations that meet its standards. Source
The policy’s approval requirement targets intermediary platforms or other services enabling third parties—other people or businesses using the service—to access OpenAI models directly. Operators must obtain review and approval before providing that access. The policy expressly says this requirement does not apply solely because an application integrates OpenAI models to power its features or functionality. Source
The distinction is between providing model access to others and using a model within an application’s features. November 20 is the revision’s effective date, separate from the October 21 announcement. The policy also gives users a route to appeal enforcement they believe is mistaken. Source
OYOPICK’s Take
For people choosing AI services, this makes the delivery arrangement relevant alongside company and model names. A service selling model access and an application offering a particular work feature have different roles to explain, even when they use the same model. Clearer descriptions of the access route and the party responsible for operational action could help customers compare capabilities and responsibilities together.
How predictable the review process becomes is worth watching. Specific criteria and clear routes for questions or resolving concerns could help developers and businesses plan around safety requirements while choosing among different ways to access models.