SIAgentNews.com
← All news
DevelopingOpenAI

OpenAI requires prior approval for AI model resellers

The AI giant is tightening control over third-party platforms that provide direct access to its models, effective November 20, 2026.

· First reported by PPC Land

1,400 wordsSize of the new policy body

The short version

  • Platforms providing direct access to OpenAI models must obtain prior review and approval starting November 20, 2026.
  • The rule does not apply to apps that integrate models to power specific features like summarizing documents or drafting emails.
  • Users are now prohibited from using OpenAI outputs to train, select, or validate data for rival AI systems.
  • The policy update expands the scope of human review requirements from automated decisions to automated actions.
  • The new policy text grew from approximately 830 words to 1,400 words in the recent rewrite.

Why it matters

This change creates a clearer legal distinction between software companies that build products and those that act as middlemen for AI access. For developers, it means a shift toward a more regulated environment where 'piping' OpenAI's technology to others requires a formal relationship with the provider.

New rules for intermediaries

OpenAI has updated its Usage Policies to require prior review and approval from operators of model-access intermediary platforms. These are services that allow third parties to tap directly into OpenAI's models, such as AI model routers or gateways. The new rule becomes effective on November 20, 2026, giving customers a 30-day notice period from the announcement.

The policy specifically targets services that provide a 'pipe' through which customers send their own prompts to an OpenAI model of their choosing. This is a significant shift because it reverses a 2022 policy where developers no longer needed to wait for approval after submitting applications. While it is narrower than the old app review regime, it restores a gatekeeping step for the reseller market.

OpenAI has not provided a specific application route or a timeline for how long the review process will take. It also did not specify which intermediaries have already been approved or how existing services will transition. The company's help center clarifies that existing agreements and express permissions will continue to apply during this transition.

Distinguishing features from access

The new policy includes a specific carve-out for applications that integrate OpenAI models to power their own features or functionality. For example, an app that uses a model to summarize a document or draft an email does not trigger the new approval requirement. This distinction helps separate software developers from those who primarily act as distributors of the underlying technology.

The language used in the policy relies on the phrase 'directly access' to separate these two types of products. A service that hands customers a direct connection to a model looks like a target for the new rule. However, a large and growing category of products, such as agent builders or workflow tools with a model picker, may fall into a gray area not explicitly covered by either side of the rule.

This distinction is important for businesses trying to understand their obligations under the new terms. Companies that build unique features on top of a model can continue as they were before. Only those who act as a gateway for others to use OpenAI's models directly will need to seek the company's sign-off before they can operate.

Restrictions on training rivals

OpenAI is also introducing a new section that prohibits using its models or outputs to prepare training material for rival AI systems. The policy states that users must not use the technology to generate, prepare, select, or supply training material for models intended as substitutes for OpenAI. This includes creating training tasks, rubrics, or reinforcement-learning environments.

The reach of this prohibition extends beyond the party actually training the final model. It applies to data labeling contractors, synthetic data vendors, and consultants who may perform the work on behalf of someone else. Even if these parties never touch the final model, they are barred from using OpenAI outputs to select or validate the training data.

OpenAI justifies this by arguing that training a model on its outputs can transfer capabilities without carrying over the necessary safeguards. While the policy does not define what constitutes a 'substitute' for its services, it could potentially reach smaller in-house models. The company notes that breaking these rules may result in the loss of access to its systems or other penalties.

Broader policy updates

The policy rewrite also includes several other updates beyond the intermediary and training rules. It rewrites the rules for child safety, renaming the section to 'Keep young people safe' and splitting it into two parts. It also widens the human review rule to apply to automated actions rather than just automated decisions.

Additionally, the update incorporates existing national security principles and sets out specific approval requirements for law enforcement and national security uses. These changes are part of a larger effort to clarify restrictions that OpenAI claims were already in place. The company describes the training language as a clarification rather than a new prohibition.

The policy also addresses a newer concern regarding 'protected reasoning' through a single sentence. While the term is not defined, it most plausibly refers to the intermediate reasoning produced by reasoning models before an answer is shown. Users are prohibited from attempting to extract this reasoning through unauthorized access or reverse engineering.

What to watch

  • How OpenAI will define 'substitute' models in practice.
  • The specific application process and review timeline for intermediaries.
  • How these rules affect the growing market of AI agent builders.

Try it

Check if your AI tool provides a 'model picker' or acts as a gateway; if it does, you may need to seek OpenAI approval by late 2026.

Sources and links

Read the original at PPC Land

First reported by PPC Land. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news