From November 20, any gateway that gives third parties direct access to OpenAI models needs OpenAI's prior approval, and using GPT outputs to train a substitute model is banned.
That leaves five days to answer two questions about your own stack. Does anyone outside your company reach OpenAI through a gateway you run? And does any team use GPT outputs to build training data, rubrics or feedback for another model? OpenAI revised its Usage Policies on October 21, 2026, with an effective date of November 20, and as of today it has published no application form, no approval criteria and no review timeline. openai.com returns 403 to automated fetches, so the policy wording below is as quoted by GovInfoSecurity, Crypto Briefing and eWeek. Check it against the live policy page before you rely on a clause.
What OpenAI Changed on October 21
OpenAI added two rules that matter to enterprise buyers: an approval requirement for intermediaries and a much wider ban on using its outputs to train competing models. Anyone running a "model-access intermediary platform or other service that enables third parties to directly access OpenAI models" must now "obtain our prior review and approval," according to GovInfoSecurity's reading of the revised text. Crypto Briefing describes the target as model routers and AI gateways that offer third-party access, and frames the line as the difference between a customer that uses a model and a reseller that redistributes access to it.
The same refresh also touched protections for minors, high-stakes uses and national-security requirements, per Crypto Briefing. For most enterprise teams those sections change less than the two above.
Two details soften the deadline. eWeek reports that existing contractual permissions remain applicable and that the earlier policies stay in force until November 20. If your organization negotiated reseller or redistribution rights into an OpenAI order form, that paper still counts.
Which Gateways Need Approval, and Which Don't?
The rule turns on who your gateway's users are, and the product you run it on doesn't change the answer. The policy "does not apply solely because an application integrates OpenAI models to power its features," per the text quoted by LLM Matchmaker. A customer-support assistant built on GPT is a feature. An endpoint that lets someone outside your company send arbitrary prompts to a GPT model is access.
An AI gateway is a proxy that sits between applications and model providers, holding the API keys and handling routing, logging, budgets and failover. Teams run them on LiteLLM, Kong AI Gateway, Cloudflare or a hosted service such as OpenRouter or Vercel's AI Gateway. LLM Matchmaker names OpenRouter, Vercel's AI Gateway and LiteLLM-style proxies as the category the rule targets, and found no public response from OpenRouter or Vercel. eWeek also found no response from OpenAI or any gateway provider.
Sorting your own exposure comes down to a few cases:
- A gateway that only your employees use looks like use. The trigger is access for "third parties," and the Cloud Security Alliance's daily brief flags that the policy doesn't say whether a gateway serving one organisation's own staff is in scope. The likely reading is that it is out of scope. OpenAI hasn't confirmed it.
- A gateway shared with subsidiaries, joint ventures or contractors is the grey zone. Whether a majority-owned affiliate counts as a third party depends on how your enterprise agreement defines "you," so read that definition before November 20.
- A product that exposes a raw model endpoint to customers (bring-your-own-prompt APIs, an "AI playground" tier, a developer platform that passes requests through to GPT) is the case the rule was written for.
- A hosted router you buy from shifts the question to your vendor. Crypto Briefing warns that developers reaching OpenAI through a third-party gateway could see access disrupted if that gateway lacks approval by the deadline.
Cloud resellers sit outside this list for now. LLM Matchmaker notes the policy doesn't say whether cloud partners such as Azure count as intermediaries. If you buy GPT through Microsoft, your contract is with Microsoft, and the purchase route already shapes your terms.
Why OpenAI Wants a Say Over Gateways
OpenAI's stated reason is enforcement: a safety block means little if a banned user can switch to another service that resells the same model. Its October post says safeguards must be effective "whether people access our models directly or through a partner," and asks hosting and reseller partners to keep safeguards consistent with OpenAI's, investigate suspected violations and coordinate enforcement, per TechJack Solutions' summary. The same summary says OpenAI is building SafetyIDs, pseudonymous end-user identifiers for abuse investigations, and that they are not generally available yet.
The argument is fair. When hundreds of tenants share one API key, OpenAI can't tell which of them sent an abusive prompt, and Crypto Briefing notes that gateways put the end user one step removed from OpenAI.
The commercial reading is harder to dismiss. An essay on DEV Community argues that routers send routine, high-volume traffic to cheaper models and forward only the hard prompts, leaving the frontier lab with the expensive peak load, and that "contractual approval gates operate as economic tolls" that could carry minimum volumes or bans on cross-provider fallback. Nothing published so far shows OpenAI attaching those conditions. Nothing published rules them out either, and an approval with no stated criteria can carry any condition OpenAI later chooses.
If you built your architecture on the router thesis that a gateway buys you failover across providers, check that the thesis still holds for the traffic you resell.
The Distillation Ban Reaches Your Own Training Pipeline
The second change affects more enterprises than the first, because it covers anyone who uses GPT outputs to build or tune a model meant to replace GPT. The policy says users cannot "use OpenAI models or their outputs to generate, prepare, select or supply training material or signals" for models intended as alternatives to or substitutes for OpenAI's, per GovInfoSecurity. Prohibited work includes creating training data, writing reinforcement learning rubrics, giving feedback and filtering data.
Distillation is the technique of training a smaller "student" model on the outputs of a larger "teacher" model instead of on a large raw dataset. Enterprises do it constantly and rarely call it that: a team asks GPT to label 50,000 support tickets, then fine-tunes an open-weight classifier on the labels so it can stop paying GPT per ticket. If that classifier replaces GPT calls, it is plausibly a substitute.
Three clauses close the obvious exits. The ban covers work done "yourself, for another party, or through someone acting on your behalf," per GovInfoSecurity, so a data-labelling vendor that uses GPT to produce your training set puts the violation on your account. It applies even if the resulting model is never trained. And evals can turn into training data: LLM Matchmaker reads the text as covering rubrics, RL environments, labels, rankings, rewards and evaluation outputs later used for training. An LLM-as-judge pipeline whose GPT scores feed a reward model is in scope.
Research, software development and model evaluation stay permitted within set boundaries. OpenAI's stated rationale is that training another model on its outputs "can transfer capabilities without carrying those safeguards over." The backdrop is CISA's warning about Chinese labs it alleges illicitly distilled OpenAI and Anthropic models, including DeepSeek and Moonshot AI, a list we covered when the advisory came out. Distillation has defenders: Nvidia's Jensen Huang and Meta's Mark Zuckerberg have argued it keeps American open-weight models competitive. The policy is written broadly enough to catch an enterprise cost-saving project along with a rival lab.
What Losing Access Looks Like
The penalty for a violation is a revoked key, and a gateway puts every team's traffic behind a few keys. OpenAI's safety-checks guide says blocks are meant to hit individual end users, but "repeated policy violations from your organization can lead to losing access for your entire organization." It also says OpenAI currently cannot unblock an individual identifier once it is blocked.
The same guide gives you the tool that keeps one user's behaviour from becoming the whole company's problem: a safety_identifier on each request, hashed so it carries no personal data. A gateway that sends one shared identity for every caller gives OpenAI nothing to block except the organization.
Gateways also concentrate keys for attackers. LiteLLM's own advisory says a malicious v1.107.0 shipped to PyPI and Docker on October 21, the same day OpenAI revised its policy, and told affected users to rotate every provider key in the proxy database. We covered what that compromise exposed. A key that leaks and gets used by someone else for third-party access or distillation is a policy problem as well as a security one.
What to Do Before November 20
This Week:
- List every consumer of your OpenAI keys. Pull the virtual-key or tenant list from your gateway and mark each one as an employee team, an affiliate, a contractor or a customer. Anything in the last two columns is a candidate for the approval rule.
- Ask your hosted router for its approval status in writing. Use eWeek's three questions: does the requirement apply to them, what authorization do they hold, and could review interrupt your access. Put a direct OpenAI key on standby as fallback for anything production-critical.
- Find your distillation jobs. Ask each ML team one question: does any GPT output end up in a training set, a reward model, a rubric or a data filter for a model that replaces a GPT call? Include work your labelling vendors do for you.
Before November 20:
- Read the "you" definition in your OpenAI agreement with legal, and check whether it covers subsidiaries. Existing contractual permissions still apply, so a negotiated clause is your strongest position.
- Send a hashed
safety_identifierper end user from the gateway, following OpenAI's guide, so a block lands on one user and not your organization. - If you expose raw model access to customers, write to your OpenAI account team now and ask how to request review. Keep the email; with no published process, a dated request is the only evidence you tried.
Before Renewal:
- Negotiate redistribution and affiliate rights explicitly, and ask for a carve-out that lets you fine-tune internal models on outputs you paid for. Under the new text, that work can count as building a substitute whenever the tuned model replaces GPT calls.
- Re-run the build-versus-route decision with the approval risk priced in. A self-hosted gateway that serves only your own staff carries the least exposure under the new text.
The Bigger Picture
Model vendors are moving from "pay and call the API" to rules about who is downstream of the call. Anthropic also updated its usage policy, banning deceptive use and tightening weapons limits, and added the human-review requirements we covered in October. In both cases the binding text is a policy page the vendor revises on its own schedule. OpenAI gave 30 days between publishing this revision and enforcing it.
For the next five days the practical question is narrower: which of your keys serve someone who isn't your employee, and which GPT outputs end up in a training set. Get both lists on paper before November 20.
Continue Reading
- LiteLLM's October 21 Compromise Puts Provider Keys at Risk
- Best LLM Gateways for Cost Control: Self-Host First
- Model Router Buyer's Guide: Buy Failover, Not Judgment
- CISA Named 6 Distillers. Your Agent Fleet Fits the Spec.
- Claude Screening of Candidates Needs Review Under Anthropic Policy
- Azure OpenAI vs OpenAI API vs Bedrock: Where to Buy GPT-6
- Stripe Bought OpenRouter. A Toggle Is Not a Contract.
