On November 20, 2026, OpenAI’s updated Usage Policies take effect — and one clause in them changes who is allowed to sell you access to GPT models. From that date, any service that lets third parties directly reach OpenAI’s models needs OpenAI’s prior review and approval first. If your app simply uses GPT inside its own features, the rule does not touch you. But if your traffic reaches OpenAI through a router, a gateway, or a proxy — OpenRouter, Vercel’s AI Gateway, a self-hosted LiteLLM setup — you have two days left to figure out where you stand, and OpenAI has still not published a form to apply on, criteria for approval, a review timeline, or a list of approved gateways.
That last part is the story. The rule itself is clear. Everything around it — how you comply, what happens if your provider hasn’t been approved, whether access gets cut off on Friday — is unpublished. This is the developer action piece: what the rule says in plain English, who it actually affects, the second rule buried in the same update that matters just as much, and the checklist to run before November 20.
What the rule actually says
OpenAI revised its Usage Policies on October 21, 2026, with the changes taking effect November 20. The new clause sits in a section called “Providing access to our models,” and the wording, as quoted by multiple outlets from the policy page, is:
“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.”
In its companion blog post from the same day, titled “Broad access through consistent standards,” OpenAI put it in plainer terms: the update means “requiring services that provide direct access to our models such as model routers or AI gateways to get our review and approval before offering that access.” The full policy text is on OpenAI’s Usage Policies page, and this detailed breakdown of the October 21 changes confirms that, as of November 17, no application form, criteria, or approved-gateway list had been published.
The policy draws one bright line, and it’s the line most developers should read twice. The approval requirement “does not apply solely because an application integrates OpenAI models to power its features or functionality.” In other words, the test is not whether you use GPT. It’s whether you hand someone else a door to GPT.
That distinction is doing a lot of work, so let me make it concrete. Your app calls the OpenAI API to summarize a document for your own user, and the user never touches the model directly — that is feature integration, and the policy says the rule does not apply to you. Your service sells API access to GPT models alongside Anthropic, Google, and Meta models through a single key, and your customers pick the model and send their own prompts — that is a model-access intermediary, and you now need OpenAI’s approval. The first case is most apps. The second case is a real and growing industry.
Who the rule targets
OpenAI names no company in the policy. The category it describes, though, maps cleanly onto a handful of well-known services. OpenRouter and Vercel’s AI Gateway are the two names every report on the change reaches for first, and for good reason: as of late October, OpenRouter listed 104 OpenAI models among the roughly 469 models on its platform, and Vercel’s gateway listed 107 OpenAI models among roughly 486. OpenRouter says it serves more than 10 million developers and companies, and Hugging Face’s own blog puts its volume at more than 800 trillion tokens a month. These are not niche tools. They are the plumbing a large share of the AI developer ecosystem now runs on.
The same category also catches self-hosted setups. Teams run gateways on LiteLLM, Kong AI Gateway, Cloudflare, Portkey, Helicone, and similar tools — software that holds the API keys, handles routing, logging, budgets, and failover between providers. The software itself isn’t the issue; the question is who sits on the other side of it. A LiteLLM proxy your own engineers use to route your own company’s traffic is one thing. The same proxy resold to outside customers as a multi-model API is another.
And then there is the case OpenAI has not addressed at all: cloud partners. Azure OpenAI, AWS Bedrock, and similar managed offerings also put OpenAI models in front of third parties, but the policy says nothing about them, and no report I found treats them as covered. Don’t assume your cloud route is safe or exposed — it is simply unaddressed, which is its own kind of risk to track.
The part nobody has answered: how do you get approved?
Here is the honest state of play as of today, November 18, two days before the rule takes effect. OpenAI has published no application form. No approval criteria. No review timeline. No list of approved gateways. No grandfathering language for services already offering OpenAI models. Nothing in the policy says what happens to a service that is still unapproved on November 20.
Neither has anyone on the other side said anything. As of November 17, no public response had come from OpenRouter or Vercel, and eWeek reported finding no response from OpenAI or any gateway provider either. That silence is worth sitting with for a moment, because it means every developer routing through a gateway is heading into a deadline with neither the platform nor the intermediary willing to say, on the record, what happens next.
One fact does soften the edge slightly. eWeek reports that existing contractual permissions remain applicable and that the earlier policies stay in force until November 20. And there is a clock that explains the date itself: OpenAI’s business contract gives customers at least thirty days’ notice when an update “materially impacts” them. October 21 to November 20 is exactly thirty days. The deadline is not arbitrary — it is the minimum notice OpenAI’s own contract required. But thirty days of notice for a rule with no published way to comply is notice of a problem, not notice of a process.
So what actually happens on Friday to a gateway that hasn’t been approved? The policy doesn’t say. The realistic possibilities range from nothing visible — enforcement arriving later, quietly, service by service — to OpenAI cutting off API keys it identifies as intermediary traffic. I am not going to pretend I know which one it is, because nobody publishing on this has said so, and the companies involved aren’t talking. What I can tell you is what to do about the uncertainty, and that is the checklist at the end of this article.
The second rule in the same update matters just as much
The gateway clause got the headlines, but the October 21 update widened a second rule that affects far more developers than the first one: the ban on using OpenAI’s models or outputs to train competing models.
The old version of the ban was already broad. The new version spells out the machinery. You may not use OpenAI models or outputs to create training data, tasks, rubrics, reinforcement-learning environments, or labels, scores, rankings, preferences, and rewards for models meant to substitute for OpenAI’s. The ban applies even if you never train a model yourself — creating the training material is enough — and it covers evaluation outputs that are later used for training. In other words, the pipeline where you use GPT to grade or label data, then feed those grades into your own model’s training run, is now explicitly named as a violation.
This is not a hypothetical workflow. It is one of the most common ways small teams build specialized models: use a frontier model as a judge, generate synthetic training data, distill a smaller model from the outputs. If any part of your pipeline does that with OpenAI outputs, the November 20 update is a deadline for you too — and unlike the gateway rule, there is no approval to apply for here. The activity is banned outright, not gated behind a review.
The same update also added national security, law enforcement, and intelligence purposes to the list of uses that need OpenAI’s review and approval, and it introduced SafetyIDs — pseudonymous end-user IDs that let OpenAI act against one abusive user without learning who that user is.
Are you affected? The decision flow
Work through these questions in order. Your answer to the first one decides most of it.
1. Does anyone besides you and your own team reach OpenAI models through your service? If the answer is no — your app calls the API, your users get the feature, nobody gets raw model access — the rule does not apply to you. The policy’s own carve-out covers you. Move on to the training-data question in section 5, because that rule has no carve-out.
2. Do you buy access from a hosted router or gateway? If you call OpenAI through OpenRouter, Vercel’s AI Gateway, or a similar hosted service, the approval question sits with your provider, not with you — but the outage risk sits with you. Your job before November 20 is to get your provider’s approval status in writing, and to have a fallback ready if the answer is silence. The checklist below walks through both.
3. Do you run your own gateway or proxy? If you operate a LiteLLM-style proxy, a Kong gateway, or your own routing layer, and outside customers or users reach OpenAI models through it, the rule applies to you directly. The grey zone is a gateway shared with subsidiaries or contractors, and a self-hosted LiteLLM instance used only by your own staff. The policy does not draw those lines, and OpenAI has not clarified them. My read: a gateway used only by your own employees is likely out of scope — the rule is about enabling third parties — but that is my reading, not OpenAI’s confirmation, and you should treat it as unconfirmed until the company says so.
4. Do you resell or white-label model access? If your product is, in effect, a storefront for model access — customers bring their own prompts, pick a model, and you pass the call through — you are the exact service the rule describes. Escalate to legal before November 20, and contact your OpenAI account team directly if you have one. This is the highest-risk category, and it is also the category with the least published guidance.
5. Do you use OpenAI outputs to train, label, or evaluate? Separate from the gateway question entirely: if your pipeline uses GPT outputs as training data, rubrics, RL environments, or labels for a substitute model, the widened distillation ban covers you from November 20. Audit that pipeline now, because there is no approval path for this one — it is a prohibition, not a permit.
The Hugging Face wrinkle: a major router switch, four days before the deadline
The timing of this rule collided with another story this week in a way that makes the stakes concrete. On November 16 — four days before the gateway rule takes effect — Hugging Face moved its Inference Providers language-model routing to OpenRouter, replacing the custom router it wrote itself. From that day, language-model requests through Inference Providers are routed by OpenRouter across more than 80 inference hosts, with automatic failover when a provider rate-limits or errors. The integration initially covers 160 models from 38 labs, and Hugging Face says it adds no markup: customers pay the provider’s list price on their Hugging Face bill.
There is one breaking change in the switch, and it is the kind that matters for this article: language-model requests can no longer use a customer’s own provider API key. Every request is billed to the Hugging Face account or organization. Image, video, and audio generation are unaffected and still connect directly to the individual providers.
Why does this matter for the November 20 rule? Because Hugging Face just moved a large volume of developer traffic onto the single largest router in the ecosystem, days before that router’s OpenAI access becomes subject to a prior-approval rule with no published approval path. To be clear about the scale: most of that traffic is open-weight models — Hugging Face says open-weight models made up 71.7% of tokens served on OpenRouter at the end of September 2026, up from 26.8% in October 2025 — and the gateway rule covers OpenAI models, not open-weight ones. The exposure is the OpenAI slice of that traffic, not the whole flow. But the direction of travel is unmistakable: routing is consolidating onto fewer, bigger intermediaries at exactly the moment OpenAI is asserting the right to approve them.
The two-day checklist: run this before November 20
If you route through a gateway, or operate one, do these in order. None of them requires an answer from OpenAI — they are all things you control.
1. Inventory every route. List every place your code or your users reach an OpenAI model: direct API calls, hosted gateways, self-hosted proxies, tools that connect through a gateway, and any vendor that passes model calls through on your behalf. Most teams find at least one route they had forgotten about. A forgotten route is the one that breaks on Friday.
2. Ask your provider for its approval status — in writing. Email your hosted gateway provider and ask, plainly: have you obtained OpenAI’s prior review and approval for the OpenAI models you offer, and if not, what is your plan before November 20? Get the answer in writing, not on a status page. If the provider doesn’t answer, treat silence as a risk signal, not as reassurance.
3. Build a direct-OpenAI fallback and test it. Stand up a direct connection to OpenAI’s API — your own API key, or Azure OpenAI if you already have that relationship — and run a real request through it from the same code path your gateway serves. Don’t just verify the key works; verify the fallback works end to end, with the same models, the same rate limits, and the same error handling. If the gateway goes dark on Friday, you want the switch to be a config change, not a project. If your workload can move to a non-OpenAI model instead, Microsoft’s new MAI models are now live in GitHub Copilot — worth knowing about as a fallback route, though switching models is a bigger change than switching routes.
4. Check your contracts. Existing contractual permissions remain applicable until November 20, according to eWeek’s reporting. If you have a contract with OpenAI or with your gateway provider, read the clauses on policy changes, termination, and service continuity now — not after an outage. If your gateway provider’s terms let it suspend OpenAI models without notice, your fallback matters twice as much.
5. Audit your training-data pipeline. This is the distillation rule, and it has no approval path. Walk your data pipeline and flag anywhere OpenAI outputs become training data, rubrics, RL environments, labels, scores, rankings, preferences, or rewards for another model — including evaluation outputs that later feed training. If you find it, stop the flow before November 20 and find a lawful source for the data. This is general information, not legal advice, and if the pipeline is central to your product, get a lawyer’s read on it this week.
6. If you expose raw model access to customers, escalate now. If your product lets customers call OpenAI models directly through you, contact your OpenAI account team before November 20 and ask how to seek approval. There is no published form, so the account team is the only route anyone has identified. Document the outreach: date, who you spoke to, what they said. If there is no account team, email OpenAI’s support channel and keep the record.
7. Watch for three things after Friday. Whether OpenAI publishes an application form, criteria, or a list of approved gateways — the moment it does, use it immediately. Whether OpenRouter, Vercel, or any other gateway responds publicly. And whether any service’s OpenAI models actually go dark. If nothing breaks on November 20, do not conclude the rule is dead; enforcement that arrives quietly and later is still enforcement.
What this rule is really about
Step back from the deadline for a moment, because the rule makes more sense in context. OpenAI’s blog post frames the change as consistency: the same standards applied to everyone who offers its models, with intermediaries matching OpenAI’s safeguards and coordinating on enforcement. The company says it worked with partners over the past few months, though it names none of them. And OpenAI is not the only lab tightening its terms this month — Anthropic made its own usage-policy move on November 12, a reminder that the rules governing how frontier models reach users are being rewritten across the industry at once. The logic is straightforward — if anyone can resell GPT access through a gateway, OpenAI loses visibility into who is using its models and for what, and it loses the ability to enforce its own policies at the point of use. A gateway that enforces nothing is a hole in every policy OpenAI writes.
Neither reading requires you to believe the rule is hostile or benign. What matters for your work is simpler: from November 20, offering direct access to OpenAI’s models through an intermediary is a permissioned activity, and the permission has no published way to obtain it yet. That combination — a real gate with no visible door — is the whole problem, and the checklist above is the practical answer until OpenAI publishes one.
If you run a gateway, you have two days to get your provider’s status in writing and your fallback tested. If you use one, you have two days to find out whether your provider has an answer. And if you use OpenAI’s outputs to train anything, you have two days to stop — because that rule has no door at all.
Frequently Asked Questions
Does the November 20 rule affect my app if I just call the OpenAI API directly?
No. If your app calls OpenAI’s API with your own key to power its own features, the rule does not apply to you. The policy says the approval requirement “does not apply solely because an application integrates OpenAI models to power its features or functionality.” The rule targets services that let third parties directly access OpenAI models — routers, gateways, and resellers.
I use OpenRouter. Will my OpenAI calls stop working on November 20?
Nobody can say for certain, and that is the honest answer. OpenAI has published no enforcement plan, and OpenRouter has made no public statement about its approval status as of November 17. Your calls may keep working, or OpenAI may cut off unapproved intermediary traffic. The safe move is to ask OpenRouter for its approval status in writing and to have a direct-OpenAI fallback tested before Friday.
Is there an application form to get a gateway approved?
Not as of November 18. OpenAI has published no application form, no approval criteria, no review timeline, and no list of approved gateways. If you operate a gateway, your only identified route is to contact your OpenAI account team directly and document the outreach.
Does the rule apply to a LiteLLM proxy I run for my own team?
Probably not, but OpenAI has not confirmed it. The rule covers services that enable third parties to directly access OpenAI models, so a proxy used only by your own employees is likely out of scope. The grey zone is a gateway shared with subsidiaries, contractors, or customers. Treat the internal-only reading as unconfirmed until OpenAI says so.
What changed about training on OpenAI outputs?
The October 21 update widened the ban on using OpenAI models or outputs to build competing models. You may not use outputs to create training data, tasks, rubrics, reinforcement-learning environments, or labels, scores, rankings, preferences, and rewards for a substitute model — and the ban applies even if you never train a model yourself, including evaluation outputs that are later used for training. Unlike the gateway rule, there is no approval path for this: it is a prohibition.
What should I do first if I only have an hour?
Inventory every route your code reaches OpenAI models through, then test a direct-API fallback from the same code path. Those two steps cover the most likely failure — a gateway going dark on November 20 — and they are entirely in your control. Ask your provider for its approval status in writing next, and audit your training-data pipeline after that.
