Did OpenAI Just Become AI's Unelected Regulator?

OpenAI did not call this a safety framework. It called it a model guide, published October 2, and that word choice is the whole story.

The guide tells you to update your AGENTS.md file to explain when documents and tests are relevant, and explicitly authorize safe routine workflows, replacing blanket approval rules with clear boundaries. That is not prompt engineering advice. That is a governance template, and OpenAI is handing it to every developer building on GPT-6 before a single regulator has weighed in on what agentic AI should be allowed to do unsupervised.

OpenAI's own docs reinforce it: because GPT-6 Astra is sensitive to files like AGENTS.md, the company says to strongly recommend auditing skills accessible to your model for instructions that could influence its behavior. Translation: the model will do what the file tells it to, so you'd better make sure the file says the right things. OpenAI gets to define what "right" means, in a document it controls, on its timeline, outside any comment period.

The Practical Label Is the Strategy

Call a document "safety guidelines" and people treat it like policy, something to debate and push back on. Call it a "practical guide" and people treat it like a manual, something to follow because it's helpful and free.

Developer experience lead Eric Provencher put it plainly: "Models have gotten much better at understanding nuance and ambiguity, so overly specific guidance can now hinder results where it previously helped."

That's a real engineering insight. It's also a company telling the market how much oversight is now appropriate, dressed as a tip.

You do not need a safety team's blessing to publish behavioral norms for a frontier model. You need distribution. OpenAI has it. A three-person lab publishing the same AGENTS.md advice gets ignored. OpenAI publishing it becomes the default every wrapper, startup, and enterprise IT department copies without a second thought, because arguing with the model maker's guide feels like arguing with physics.

Follow the Pricing

The guide also cements a three-tier hierarchy of trust and cost. Astra runs at $10/$50 per million input/output tokens for the hardest reasoning work. Sol, the mid-tier agentic coder, is $2 per million input tokens and $10 per million output tokens. Luna, the bulk-classification model, runs $0.10 and $0.50. Caching cuts costs further; OpenAI says cached input tokens cost up to 95% less than uncached ones.

Every number is a lever only OpenAI controls, and the guide tells you how to pull it "responsibly."

OpenAI is the only company positioned to write the etiquette manual for a technology category it also happens to sell.

What to watch: whether Anthropic, Google, or Meta publish their own "practical guides" for frontier agentic models in the next quarter. If they don't, OpenAI's AGENTS.md norms become the de facto industry standard before any regulator writes a word of binding law.