OpenAI Model Resellers Must Meet the Same Safety Standards
The revised usage policies take effect on November 20. Intermediary services that connect users directly to OpenAI models will need review and approval before they open access to users.
OpenAI announced on October 21 local time that it is revising its usage policies. The updated policies take effect on November 20, 2026. Alongside the revision, the company said it will hold the businesses that provide access to its models to the same standards it applies to itself.
The core of the policy is the responsibility of partners that supply the models. OpenAI said partners that make its models available must maintain safeguards at the same level as OpenAI’s own. Partners must investigate suspected policy violations themselves and take appropriate action when concerns surface. In other words, the same safety standards apply no matter which route a user takes to reach the models.
Do model connection services need prior approval?
Yes. The requirement covers services that give users direct access to OpenAI models. That includes model routers and AI gateways, which take in requests and pass them along to the models. These services must obtain OpenAI’s review and approval before opening up access to users.
There are implications for companies and developers that use OpenAI models through another company’s service as well. OpenAI said it is assigning safety management responsibility to the businesses that sit in the middle and open up model access. Along with the prior approval requirement, they also take on the obligation to investigate violations and follow up on them.
OpenAI said GPT-6 Astra expands the range of what developers and enterprises can build. The company said it evaluates performance and risk and tests safeguards before releasing a model. After release, it continues to monitor and investigate the potential for misuse, and it applies trusted access programs that set eligibility requirements and conditions for its most capable features.
SafetyIDs: investigating without revealing identity?
OpenAI said it is working with customers to introduce SafetyIDs, which distinguish between users during misuse investigations. When a customer supplies SafetyIDs, OpenAI can tell users apart in the course of an investigation without knowing the identity of the end user.
It is a mechanism that pairs with the investigation obligations placed on partners. The idea is to require partners to look into possible violations, while making it possible to distinguish between users as investigations require without revealing their identities.
Businesses in Korea that provide access to or resell OpenAI models should check the scope of the policy and the approval requirements in advance. With the November 20 effective date approaching, the key question is how the actual approval process and the adoption of SafetyIDs will affect day-to-day service operations.