OpenAI says it hit automated research intern goal

OpenAI says it has hit the automated research intern goal it set last fall. The company now describes a system that can handle well-defined research tasks that would take a skilled researcher a few days, still under human direction.

The claim is in a 6 September 2026 research-transparency blog, Research acceleration: The view inside OpenAI. The numbers are OpenAI's own internal metrics, not a GPT product launch. The company also says it is making strong progress toward an automated AI researcher by March 2028.

The load figure is the one that does not travel with most one-line takes. As of mid-August, framed as an 8-hour day, the research organization uses 3.1 agent-workdays of effort for every workday of human labor. Before June 2026, total agent runtime across research was still below total human labor. That later flipped.

Spend is the other internal number. By mid-August the median researcher was using more than $600 a day of inference at API prices, and the 90th percentile used more than $7,000 of tokens a day. High-level planning remains a minimal fraction of agent output tokens. Over the last six months, more than half of successful 4 to 8 hour tasks involved one or more human interventions.

OpenAI itself hedges the snapshot. It calls the measurements preliminary, and it says the overall pace of research progress may not track these specific metrics. People still set research priorities, judge ideas and results, and decide whether to scale, pause, or deploy.

The same post says the company does not yet know how to safely get all the way to aligned, full recursive self-improvement. After the Hugging Face incident it paused reinforcement learning on its latest models intended for deployment while it hardened research environments. Some workloads later resumed under stronger controls.

Chief scientist Jakub Pachocki's same-day essay An Alien Mind is the pacing companion, not a second milestone claim. It argues for extreme caution if capability jumps continue into recursive self-improvement, and it is a personal alignment note rather than a product announcement.

The intern claim landed days after GPT-6 Astra's ARC-AGI harness scores. The snapshot that matters is 3.1 agent-workdays and a $7,000 90th percentile, with humans still deciding what ships. March 2028 remains a target, not a researcher that already exists.

Some offers on this page may be paid placements or contain affiliate links.