The Model That Reads Your Screen and the Price That Fell

Computer-use used to be a demo you watched at a keynote and forgot by lunch. GPT-6.1 Terra's price cut is the moment it becomes a line item a normal team can afford, and that is a different kind of change than a benchmark beating.

The numbers matter because automation is a volume game. At around one dollar per million input tokens and five per million output, the cost of letting a model open an app, click through a form, and file the result drops low enough that you run it a thousand times a day instead of admiring it once. When the price of an agent step falls by half, the break-even point for automating a boring task moves from enterprise to solo founder.

I am wary of the hype here, because computer-use is also where agents fail loudest. A model that reads a screen and acts on it is one typo away from booking the wrong flight or emailing the wrong client. The labs know this, which is why the more capable versions are gated behind vetted programs rather than opened wide. Cheap and autonomous at the same time is exactly the combination that needs a human watching.

The real shift is not capability, it is permission. Half price does not make the model smarter; it makes the experiment affordable. The teams that benefit are the ones with a hundred small, repetitive, browser-bound tasks, not the ones waiting for a robot to think for them. Buy the cheaper step, keep the human on the irreversible actions, and the economics finally work.