OpenAI Says Its AI Now Works Like a Research Intern

OpenAI has put numbers to a claim that until now lived mostly in interviews and hallway talk: that its own models are starting to speed up the work of building its next models. In a post published this week and picked up by Help Net Security, the company said it has reached a goal set last autumn of having an automated research intern by September 2026.

The word intern is doing careful work in that sentence. OpenAI is not claiming an autonomous scientist. It describes a system that can carry out well-defined research tasks under human direction, including jobs that would take a skilled person several days. A supervisor still sets the problem, checks the output and decides what happens next. What has changed, the company says, is how much of the middle can now be handed off.

The figures

The internal data OpenAI released is the interesting part, because it is unusually specific. Over roughly six months, the share of research compute going to internal coding inference grew about a hundredfold, and agentic token usage rose around twenty-twofold. By mid-August, the company says, the median researcher was running more than $600 a day of inference on internal agents at standard API prices, and the heaviest users were burning through north of $7,000 a day.

One measure stands out. OpenAI reports that its research organization now uses 3.1 agent-workdays of effort for every workday of human labor, and that experiments per active researcher hit an all-time high in August. Read plainly, that means the agents are doing the bulk of the hours, with people steering.

What to make of it

A note of caution belongs here. This is self-reported data from a company with an obvious interest in looking fast, and none of it has been independently audited. "Agent-workdays" is OpenAI's own yardstick, not an external benchmark, and a dollar figure for internal inference says more about volume than about the quality of what came out. The honest read is that agents are handling a great deal of routine engineering and experiment-running, which is real, without settling the harder question of whether they are producing genuinely new ideas.

Still, the direction is worth sitting with. The pattern of AI helping to build AI is exactly what people mean when they worry, or hope, that progress could compound. It is a quieter version of the story than a dramatic model launch, and possibly a more consequential one. We have already seen smaller signs of it, including a compact research agent that outdid larger models at reproducing published science.

OpenAI framed the intern as a waypoint, not a destination. It says the next target is an automated AI researcher, capable of tackling open-ended problems rather than assigned ones, by March 2028. That is a much taller claim, and the company will have to show it rather than assert it. For now, the milestone that matters is smaller and already here: the machines are doing the homework, and the researchers are grading it. The company's own GPT-6 Astra, released last week, is the kind of model that work is meant to accelerate.