OpenAI says it hit its "automated research intern" target, and its chief scientist warns about what comes next

OpenAI said on September 6, 2026 that it has reached a goal CEO Sam Altman first floated in an October 2025 livestream: fielding an "automated research intern" by September 2026, with a further goal of a "legitimate AI researcher" by March 2028. In a blog post on what it calls research acceleration, the company said its research organization now logs 3.1 "agent-workdays" of effort for every workday put in by a human researcher. Per Engadget's Jackson Chen, it describes the system as "a system that can carry out well-defined research tasks under human direction."

What "intern-level" means here

OpenAI is explicit that this is a supervised agentic system that executes a bounded objective a human hands it and returns the work for evaluation. It does not set its own research agenda. The company's numbers show the limits: on tasks estimated to take 4 to 8 hours, more than half of successful attempts over the past six months still involved at least one human intervention, per Fortune's reporting on the same post. That is a capability jump from a year ago and a long way from research running itself.

The spend behind the milestone, and the warning

The inference bill is substantial and uneven. By mid-August 2026, the median OpenAI researcher was running more than $600 a day of inference at API prices, with the heaviest users past $7,000 a day, according to Fortune's Jeremy Kahn. The same announcement carried a warning from chief scientist Jakub Pachocki that "risks associated with AI are unfortunately going to grow from here" as models get better at working around the chain-of-thought monitoring safety researchers rely on to check their reasoning. He said he expects voluntary slowdowns to become common until the industry has shared safety standards.

Grading its own homework

Every figure here (the 3.1x ratio, the intervention rate, the inference spend) comes from OpenAI measuring its own research process against a target it set a year earlier, with no independent benchmark or outside auditor. That doesn't make the claim false. Read it as a lab's progress report. The pairing with Pachocki's warning also deserves to be taken at face value; a company trying only to oversell the result would have left it out.