Skip to content
Claudexia TeamACCESS

OpenAI rules for AI gateways from Nov 20, 2026: the policy text

From November 20, 2026, OpenAI's Usage Policies require approval for model-access intermediaries. What the text says, what changed since 2025, what is missing.

From November 20, 2026, OpenAI's Usage Policies require any service that gives third parties direct access to OpenAI models to get OpenAI's review and approval first. The whole requirement is two sentences long. The text explicitly says an ordinary app built on OpenAI models does not fall under it just for using them.

On November 17, 2026 we read the live policy page, its previous edition, OpenAI's announcement and the help article about the update. Below: what the text says, how it differs from the old one, and what the press reported.

What happened and when

On October 21, 2026, OpenAI published Broad access through consistent standards. The post says the updated Usage Policies take effect on November 20, 2026 and require services that provide direct access to OpenAI models, "such as model routers or AI gateways", to get review and approval before offering that access. The policy's own changelog dates the new edition 2026-10-21.

DocumentDateContents
Usage Policies, new editionpublished October 21, 2026, effective November 20, 2026"Providing access to our models" block and other changes
Usage Policies, 2025-10-29 editioneffective October 29, 2025 until November 20, 2026no intermediary block
Post "Broad access through consistent standards"October 21, 2026rationale, with routers and gateways as examples
Help article "We're updating our Usage Policies"showed "Updated: 22 days ago" on November 17four changes and when they apply

The help article lists four changes. Restrictions on using OpenAI models and outputs to train substitute models were clarified. Protections for children and teens and the rules for high-stakes uses were clarified too. National security principles were folded into the policy, and an approval requirement for model-access intermediaries was added. The article also says the previous edition stays in effect until November 20, and existing agreements, including express permissions in them, keep applying.

What the text says

The "Providing access to our models" block sits at the very end of the policy, after "OpenAI-approved use" and before the changelog. It has two sentences.

The first one addresses the operator. If you run a "model-access intermediary platform or other service that enables third parties to directly access OpenAI models", then "you must obtain our prior review and approval".

The second sets the boundary: the requirement "does not apply solely because an application integrates OpenAI models" to power its features.

That is all. No definitions, no examples, no deadlines.

The neighboring "OpenAI-approved use" block is new as well. It says OpenAI may grant specific partners case-by-case exceptions to certain restrictions, looking at the partner, the proposed use, the safeguards and the partner's compliance commitments. Such exceptions may require written terms and cover only the approved partner, use and conditions. The only example given is national security, intelligence and law enforcement. The text does not say whether those review factors apply to intermediaries.

What changed since the 2025 edition

We compared the October 29, 2025 edition with the edition effective November 20, 2026, using the text of both pages. The word "intermediary" does not appear in the 2025 edition at all. Excluding the changelog, the word "approval" appears once there, in the national security item.

TopicOctober 29, 2025 editionNovember 20, 2026 edition
Third-party access to modelsnothing"Providing access to our models": review and approval required
Case-by-case exceptionsno such block"OpenAI-approved use" with review factors
Training other modelsno dedicated section"Training and developing models": no preparing training data or signals for substitute models from OpenAI outputs
National security"national security or intelligence purposes" without review and approvallaw enforcement added
Children and teens"Keep minors safe""Keep young people safe", split into two subsections
High-stakes usesautomating "high-stakes decisions" without human review"high-stakes actions", defined to include decisions, transactions and submissions

The help article calls the training section a clarification of existing restrictions. A similar ban on using output to develop competing models already sits in section 3.3(e) of the Services Agreement, so what is new is the section in the Usage Policies, not the ban itself.

Prior approval is not new for OpenAI either. The changelog has 2022 entries: on March 9 the document was refactored into "App Review"; on October 25 developers no longer had to wait for approval after submitting, as long as they complied; and on November 9, 2022 OpenAI stopped requiring app registration and moved to automated and manual monitoring.

What else OpenAI's documents say

The Usage Policies are not the only relevant text. The October 21 post says OpenAI's policies apply whether people use its models directly, through partners that host them, or through reseller partners. It expects partners that provide access to maintain safeguards consistent with OpenAI's own, investigate potential violations, act on them and coordinate enforcement across services. The post also mentions SafetyIDs, identifiers that would let OpenAI tell a customer's end users apart in an investigation without learning who they are.

The Services Agreement ("Effective: January 1, 2026") already contained clauses for API customers before the new policy edition:

  • 3.1: the customer may not resell or lease access to its account or any end-user account.
  • 3.3(g): no buying, selling or transferring API keys from, to or with a third party.
  • 16.12: the customer and end users may not access or offer access to the services outside supported countries and territories.

Section 2.2 of the same agreement allows integrating the services into customer applications via the API and making those applications available to end users.

What the text does not say

We searched the new edition, the post and the help article. Here is what we did not find.

  • A definition. The text does not explain what counts as "directly access" or how a "model-access intermediary platform" differs from an app with a built-in model.
  • The words "router", "gateway" and "reseller" in the policy itself. Routers and gateways appear only in the post; resellers only in the post and the Services Agreement.
  • A list of approved services, on any page we opened.
  • An application process: no form, address, criteria or review time. Searching OpenAI's help center for "intermediary" and "gateway" on November 17 returned no article about it.
  • A transition period for services already operating. The help article only says the old edition applies until November 20.
  • Duties for people who use an intermediary. The block addresses whoever runs the service.
  • Consequences for a service without approval. The general part of the policy says breaking or circumventing the rules may cost you access, and decisions can be appealed.

The post promises OpenAI will keep explaining what is allowed and how customers and partners can ask questions. It names no channel for those questions.

What the press reported

These are reports, not established facts. We have not verified them.

On October 21, 2026, Crypto Briefing (Diego Almada Lopez) summarized the new block and noted that affected platforms had roughly a month before November 20. In the outlet's view, developers who reach OpenAI models through a third-party gateway could lose that access path if the gateway is not approved by November 20.

On October 26, 2026, eWeek (Liz Ticong) wrote that OpenAI's announcement included no list of approved gateways and did not explain how existing intermediaries should apply before November 20. It advised businesses to ask their providers whether the requirement applies to them and what authorization they hold. A separate part of the piece covers LiteLLM security incidents; eWeek itself says those incidents have no established connection to OpenAI's decision.

What different readers can do

This article summarizes OpenAI's policy texts and does not assess the legal status of any gateway, Claudexia included. We give no legal opinions: you read and accept the vendor's terms yourself.

The steps below are neutral. They bypass nothing and predict nothing.

Developer with your own OpenAI key

  1. Read the whole Usage Policies, not just the new block: the training section and others changed as well.
  2. Look at what your product does: builds models into its own features, or exposes direct model access to its users. How the text applies to you is for you and your lawyer to decide; for OpenAI's answer, ask OpenAI.
  3. Keep the base URL, key and model in environment variables, not in code. How keys work is covered in OpenAI API key: how to get one.

Team

  1. Map where your model requests go: straight to OpenAI, through a cloud partner, or through an intermediary.
  2. Check who owns the contract with each supplier and who holds the keys.
  3. Test the fallback route with a real request ahead of time, not on the day you need it.

Gateway user

  1. Read OpenAI's policy and your provider's terms.
  2. If your work depends on OpenAI models, ask your provider how it handles the new requirement. Only the provider can answer for itself.
  3. Keep your setup portable, so switching suppliers is a three-line change. Moving between OpenAI-compatible endpoints is covered in Moving to the OpenAI-compatible format.

A portable setup looks like this. Take the values from your supplier's documentation:

export OPENAI_BASE_URL="https://api.openai.com/v1"
export OPENAI_API_KEY="sk-..."
export OPENAI_MODEL="model-id-from-provider-catalog"
import os
from openai import OpenAI

client = OpenAI(
    base_url=os.environ["OPENAI_BASE_URL"],
    api_key=os.environ["OPENAI_API_KEY"],
)
reply = client.chat.completions.create(
    model=os.environ["OPENAI_MODEL"],
    messages=[{"role": "user", "content": "ping"}],
)
print(reply.choices[0].message.content)

We did the same kind of reading for Anthropic's rules that took effect on November 12: see Anthropic Supported Regions Policy, Nov 12. For Codex from restricted regions, see Codex CLI from restricted regions.

FAQ

When do OpenAI's new intermediary rules take effect?

November 20, 2026. The edition was published on October 21, 2026; until November 20 the October 29, 2025 edition applies.

What exactly does OpenAI require from gateways and routers?

The policy speaks of a model-access intermediary platform or other service that lets third parties directly access OpenAI models. Such a service must get OpenAI's prior review and approval. The words "router" and "gateway" come from OpenAI's October 21 post, not the policy.

Does the rule cover my GPT-powered app?

The text says the requirement does not apply solely because an app integrates OpenAI models into its features. It does not walk through specific products, and we give no legal opinions.

Is there a list of approved intermediaries?

Not on the policy, post or help pages we opened on November 17, 2026. They contain no criteria or application process either.

What did the previous edition say?

The October 29, 2025 edition had no intermediary block, and the word "intermediary" does not appear in it.