OpenAI has rewritten the rules for who can sell access to its models and what anyone may do with their outputs. The company's updated Usage Policies, published on 21 October and effective from 20 November 2026, require AI gateways and model routers to obtain OpenAI's prior review and approval before offering direct access to its models. The same update bans using OpenAI models or their outputs to train rival systems, drawing the firm's clearest line yet against distillation.
The gateway rule
The rule catches any service that lets third parties reach OpenAI models directly. "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," the policy says. The requirement does not apply merely because an application integrates OpenAI models to power its own features.
Partners that provide access must also keep safeguards consistent with OpenAI's own, investigate potential violations, and take action when concerns arise. That includes coordinating enforcement so a user blocked on one service cannot simply switch channels and keep going. OpenAI frames the change as consistent standards across every route to its models.
Neither the blog post nor the policy page explains how an intermediary should seek that approval, and no application route is described. The gap matters because the rule bites in twelve days: any gateway reselling direct access without approval would be in breach from 20 November, whether or not it has heard back.
The distillation ban
The second change is a new "Training and developing models" section, and it is written without loopholes. OpenAI bars using its models or their outputs to generate, prepare, select or supply training material or signals for models intended as alternatives to or substitutes for its own services.
That covers creating training data, tasks, rubrics and reinforcement-learning environments, and supplying feedback used in training, from labels and scores to rankings, preferences, rewards and corrections. It also bans selecting, filtering or validating training data, tasks or trajectories with OpenAI outputs.
The ban also covers attempting to extract protected reasoning through unauthorised access, reverse engineering or circumvention of safeguards. The restrictions apply whether the work is done directly, for another party, or through someone acting on your behalf, even if you never train the resulting model yourself.
Research, evaluation, benchmarking and product development remain permitted within those boundaries, and uses expressly allowed by a customer's agreement with OpenAI are unaffected by the new wording. The line OpenAI draws is between learning from its models and learning to replace them — a distinction the whole industry will now have to navigate.
SafetyIDs and sensitive uses
The update also folds in approval requirements for national-security, intelligence and law-enforcement uses, and introduces SafetyIDs: identifiers customers can supply so OpenAI can tell individual end users apart during an abuse investigation without learning who they are. With them, the company says it can respond to misuse more precisely while disrupting fewer innocent users.
The blog presents all of this as one programme: safeguards that hold whether people reach OpenAI's models directly, through a partner that hosts them, or through a reseller. Enforcement, it says, will rest on documented conduct and risk, applied consistently and in proportion, with restrictions revisited when conduct or commitments change.
Our Read
The timing gives the update its edge. It lands weeks after DeepSeek and Moonshot faced accusations of learning from OpenAI's Astra model, and the distillation ban reads as the policy answer to that dispute: no more arguing about what counts as learning from a rival, because the rule now names the practice outright.
The gateway rule is the more consequential half. Routers and gateways have become the plumbing of the model market, letting developers switch providers without rewriting code. By making direct resale conditional on its approval, OpenAI gains a veto over who sits between it and the end user, and a lever over the safeguards those middlemen run.
The open question is process. A rule that demands approval without publishing how to get it leaves every gateway guessing for the next twelve days. OpenAI's own framing — clear rules, predictable enforcement — will be judged on whether that route appears before the deadline, not after it.
Why It Matters
For gateway and router operators, the deadline is concrete: approval by 20 November, or stop offering direct access to OpenAI models. For developers, the choice of where to buy model access narrows, and the price of switching providers rises if the middlemen must each clear OpenAI's review first.
For the industry, the distillation ban hardens a norm that was argued case by case. Labs have long suspected rivals of training on their outputs; now the lab has written the prohibition into its rules, with the same safety push behind it as its GPT-6 Cyber model. Other labs are tightening in parallel, and the direction of travel is unmistakable.
Background
OpenAI rewrote its Usage Policies in October 2025, when it unified them across products. This update keeps that structure and adds the two new controls, alongside clarified protections for children and high-stakes uses. The pattern matches a wider shift: Anthropic rewrote its own usage rules the same month, adding a cruelty rule for Claude that takes effect on 12 November.
Source: OpenAI's announcement and OpenAI's Usage Policies
0 Comments