TL;DR: OpenAI has published new Usage Policies that take effect on November 20, 2026. Three changes matter most for builders. From that date, any service that lets third parties access OpenAI models directly, such as a model router or AI gateway, needs OpenAI’s “prior review and approval,” though an app that simply uses OpenAI models to power its own features doesn’t. The rule against training competing models is spelled out in detail: OpenAI outputs can’t be used to create training data, rubrics, reinforcement-learning environments, labels, scores or rankings for “alternatives to or substitutes for” OpenAI models. And automating high-stakes actions in areas such as hiring, credit, housing, insurance and medical care without human review is listed as prohibited. If your product resells or routes access to OpenAI models, or you use GPT outputs anywhere in a model-training pipeline, check it against the new text before November 20.

What happened

On October 21, 2026, OpenAI updated its Usage Policies, effective November 20, 2026. The page’s changelog says the update clarifies “restrictions on using OpenAI models and outputs to train other models,” protections for children and teens and “requirements for high-stakes uses,” and adds approval requirements for national security and law enforcement uses “as well as for model-access intermediaries.” OpenAI explained the change in a post, “Broad access through consistent standards,” published at 21:00 UTC on October 21 (5:00 a.m. Manila on October 22).

What changes

  • Model access through a third party. The new policy says: “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. This requirement does not apply solely because an application integrates OpenAI models to power its features or functionality.” OpenAI’s post names “model routers or AI gateways” as examples. It also says partners that host or resell its models must “maintain safeguards consistent with our own” and coordinate on enforcement.
  • Training other models. You “must not use OpenAI models or their outputs to generate, prepare, select, or supply training material or signals for models that are intended as alternatives to or substitutes for OpenAI models or services.” That explicitly covers creating training data, tasks, rubrics or RL environments; providing labels, scores, rankings, preferences, rewards or corrections; and selecting, filtering or validating training data. It applies even if you do the work for someone else and “even if you do not train the resulting model.” Research, software development, model evaluation and benchmarking, and product development remain allowed “within these boundaries,” as do uses your agreement with OpenAI expressly permits.
  • High-stakes automation. The list of prohibited uses now includes “automation of high-stakes actions in sensitive areas without human review”: actions and decisions that materially affect a person’s rights, safety or access to essential services, in areas including critical infrastructure, education, housing, employment, financial activities and credit, insurance, legal, medical, essential government services and migration.
  • Licensed advice and credentials. Building tools that give others tailored advice requiring a license, “without appropriate involvement by a licensed professional,” stays banned, and you “may not state or imply that the AI itself holds professional credentials or authorization.”
  • Abuse handling. OpenAI says it is working with customers on “SafetyIDs,” which let a customer separate individual end users in an investigation without OpenAI learning who they are, so it can respond to potential misuse “more precisely while reducing disruption to others.”

What to do now

  1. If you run a router, gateway, reseller or “bring your own model” platform that passes OpenAI models through to your customers, contact OpenAI about review before November 20. If your app just uses GPT inside its own features, the policy says this rule doesn’t apply to you.
  2. Audit any training or fine-tuning pipeline for GPT-generated data, grading rubrics, labels or synthetic evals. If the model you’re building could be called an alternative to OpenAI’s, that use is prohibited from November 20 unless your agreement says otherwise.
  3. Look at automations that decide things about people, such as screening job applicants, approving credit or handling insurance claims. Add a human review step before the action takes effect.
  4. Check your product copy for anything that implies the AI is a licensed lawyer, doctor or adviser.

If your gateway or proxy also handles Claude traffic, Anthropic changed how its new SDKs send API keys the same day; see our story on Claude API version 2026-10-21.

Sources & method

Sources

Method. Last checked October 22, 2026, at 9:18 a.m. Manila time (01:18 UTC on October 22). The post time comes from OpenAI’s news RSS feed. Quotes are from the policy page and OpenAI’s post; the checklist is our own reading and is not legal advice. No affiliate links on this page. How we work.

Spotted an error? See our corrections policy and tell us.