The Week the Frontier Got Cheaper and More Gated at the Same Time

Same week, opposite moves. One frontier model got cheap enough for anyone to automate with; another got locked behind a vetting program. Access to the best AI is splitting in two directions, and which one you hit depends on what you are allowed to do.

GPT-6.1 Terra's price cut is the democratic half. At roughly one dollar per million input tokens and five per million output, the economics of letting a model click through a browser task flip from enterprise-only to solo-founder. Computer-use stopped being a demo and became a line item, and that is the kind of change that shows up as a hundred small automations, not one headline product.

Gemini 4 Carbon is the gated half. Topping the intelligence index at 59 and leading a security benchmark meant it was exactly the model you would want, and exactly the model Google chose to ration through a vetted-access program rather than open to the public API. The pattern is now standard: the most capable model goes to a few trusted groups first, everyone else waits.

I read the split as a signal about risk, not capability. Cheap and autonomous at the same time is the combination that needs a human watching, so the labs open the cheap one wide and hold the powerful one close. For a builder, the practical move is to build on the cheap, available layer now and treat the gated layer as something you apply for when the task truly needs it.