OpenAI will require approval for routers and AI gateways that serve its models

If you use OpenAI models through a router or AI gateway, ask that provider about approval now. From Nov 20, OpenAI's updated Usage Policies require such services to get OpenAI's sign-off first.

What changed#

OpenAI posted the update on Oct 21. The new clause reads: "If you operate a model-access intermediary platform or other service that enables third parties to directly access OpenAI models, you must obtain our prior review and approval." OpenAI's post names "model routers or AI gateways" as examples. The next sentence in the policy sets the limit: "This requirement does not apply solely because an application integrates OpenAI models to power its features or functionality."

The policy's changelog says the update also clarifies "restrictions on using OpenAI models and outputs to train other models," protections for children and teens, rules for high-stakes uses, and approval for national security and law enforcement uses. On training, you may not use outputs to "generate, prepare, select, or supply training material or signals" for models "intended as alternatives to or substitutes for OpenAI models or services," including rubrics, rankings and reinforcement-learning environments.

Why it matters#

OpenAI hasn't said what happens to a service that lacks approval on Nov 20. Our read: if you depend on a router, a direct OpenAI API key is a cheap backup to set up now. Teams that grade or filter data with GPT models should also check where that data ends up.

How to use it#

1. Ask your router or gateway provider whether it has OpenAI's approval (our suggestion).
2. If you only call the OpenAI API to power your own app's features, the policy says the rule doesn't apply "solely" for that reason.
3. Check that no workflow turns OpenAI outputs into training data, rubrics or rewards for a rival model.
4. Tag each end user with safety_identifier. Per OpenAI's docs: "Hash user email or internal user IDs to avoid passing any personal information." Realtime uses the OpenAI-Safety-Identifier header.

OpenAI's example, trimmed:

response = client.chat.completions.create(
    model="gpt-6-astra",
    messages=[{"role": "user", "content": "This is a test"}],
    max_completion_tokens=5,
    safety_identifier="user_123456",
)

Good to know#

• This is a policy summary, not legal advice. The policy doesn't define "model-access intermediary" beyond its own wording. Our read: if your product sits near the line, such as an app that lets users pick models, ask OpenAI or your counsel.
• There's no application process yet. OpenAI hasn't said how to apply, how long review takes, or which routers are approved.
• Nothing changes before Nov 20. Per the Help Center, "until then, the previous version remains in effect," and "express permissions in applicable agreements" still apply.
• Evaluation is still allowed. The policy says you "may use OpenAI models for research, software development, model evaluation and benchmarking, and product development, within these boundaries." The training limits apply "even if you do not train the resulting model."
• SafetyIDs are not explained yet. OpenAI says it is "working with customers to introduce SafetyIDs." It doesn't say whether these are the existing safety_identifier parameter, which the docs say helps OpenAI "block individual end users, rather than blocking your entire org."

What to watch: an approval process or approved-provider list from OpenAI before Nov 20.