OpenAI's capability timeline and its regulatory ask, in one essay

The metrics OpenAI chose to put on record

This is a vision essay, but it leans on a handful of specific numbers, and those are worth separating from the rhetoric. OpenAI says AI in software engineering has moved from tasks a person could do in a few seconds to tasks that take a person more than an hour, and expects systems soon that handle work taking days or weeks.

It also claims the cost per unit of a given level of intelligence has fallen roughly 40x per year over the last few years. And it puts dates on discovery: very small discoveries in 2026, more significant ones in 2028 and beyond.

These are the falsifiable parts. A task-horizon figure and a cost-decline rate can be checked against products people actually ship. The 2026 and 2028 discovery milestones are hedged—OpenAI adds it 'could of course be wrong'—but they are still on the record with dates attached.

The '80%' framing and the gap it points to

The essay's most quotable line is a self-assessment: systems that can solve very hard problems 'seem more like 80% of the way to an AI researcher than 20% of the way.' It's a striking claim precisely because it is unquantified—there is no benchmark behind the percentage, just an impression.

More useful is the observation OpenAI keeps returning to: the gap between how most people use AI and what it can currently do is 'immense.' Most of the world, the company writes, still thinks of AI as chatbots and better search.

For anyone building products, that gap is the actual message. The frontier of capability and the frontier of adoption are far apart, and OpenAI is arguing the bottleneck is not the model.

A regulatory proposal split into two tracks

The recommendations divide the world into two scenarios. In the 'normal technology' track, AI progresses like the printing press or the internet, and conventional policy tools suffice. In that case OpenAI wants today's capability level to 'diffuse everywhere' with minimal additional regulatory burden—and explicitly rejects a '50-state patchwork.'

The second track is superintelligence developing at a speed 'humanity has not seen before,' where OpenAI argues typical regulation won't do much either, and coordination shifts to executive branches and safety institutes across multiple countries.

The structure is worth noting: light-touch treatment for what exists now, extraordinary coordination for what might come. The essay raises the possibility that the field should collectively slow development as systems approach recursive self-improvement, but frames that as a decision informed by empirical safety research rather than a commitment.

AI as utility: the claim that carries the most weight

The closing recommendation is the one with the largest practical implications for how OpenAI positions its products.

We expect access to advanced AI to be a foundational utility in the coming years—on par with electricity, clean water, or food.Montana Labs

Framing AI access as a utility does specific work. It supports the argument for minimal regulation on current-generation deployment, it recasts wide distribution as a public good rather than a commercial goal, and it sets up 'individual empowerment' as the north star—adults using AI on their own terms within bounds set by society.

The tension in the essay is that the same document arguing today's AI should diffuse like a utility also predicts systems that could make scientific discoveries within years. A team taking these claims seriously has to plan for both realities at once: build on infrastructure OpenAI wants treated as basic, while the vendor itself flags catastrophic risk at the far end. Which of those two framings governs a given deployment is left, deliberately, to the reader.