OpenAI's new usage policies put model routers and AI gateways behind prior approval and spell out a ban on using its outputs to build RL environments, rubrics and reward data for rival models; effective Nov 20

Why it matters

【Key points】

On Oct 21 OpenAI updated its Usage Policies, effective Nov 20, 2026.

1. Intermediaries need approval: any model-access intermediary or service that lets third parties directly access OpenAI models (the post names model routers and AI gateways) needs OpenAI's prior review and approval. Apps that merely integrate OpenAI models are exempt. Hosts and resellers must match OpenAI's safeguards and coordinate enforcement.

2. Anti-distillation, spelled out: no using OpenAI models or outputs to create training data, tasks, rubrics, RL environments, or labels, scores, rankings, preferences and rewards for models meant to substitute for OpenAI's. Applies to work done for others too; no extracting protected reasoning. Research, evaluation and benchmarking stay allowed.

3. SafetyIDs: customers can pass pseudonymous end-user IDs so OpenAI can act on one user without learning who they are.

4. National-security and law-enforcement uses now need approval.

【Verifiable】

https://openai.com/index/broad-access-through-consistent-standards

https://openai.com/policies/usage-policies/

【Our read】

1. Background: OpenCode blamed a "large-scale distillation attack" for leaked OpenAI, Anthropic and Google keys, and DHS ordered DeepSeek removed amid distillation allegations. OpenAI now extends its old "no competing models" term to RL environments, rubrics and reward signals, which reaches synthetic-data vendors.

2. Comparison: Anthropic's API user profiles (Oct 20) and Perplexity's new Safety Identifier header take the same approach: penalize the end user, not the whole org. OpenAI goes further by putting routers and gateways behind prior approval.

3. Impact: OpenRouter- and LiteLLM-style gateways and cloud resellers face a licensing regime, and teams that use frontier outputs for training face clearer legal risk. Enforcement still relies on partners.

4. Caveat: approval criteria and transition terms are unpublished.