← Back to feed

OpenAI published a transparency post on September 6 confirming it has reached the “automated research intern” milestone announced by Sam Altman last fall: AI systems are now performing 3.1x the research effort of its human staff.

The 3.1x figure measures total AI-contributed research work relative to what human researchers alone would produce. OpenAI defines an “automated research intern” as a system capable of carrying out well-defined research tasks under human direction, including tasks that would take a skilled researcher several days. The threshold has now been met on schedule.

What Changed Inside OpenAI

Researchers are running coding agents throughout the day, often in concurrent sessions. Usage is growing faster inside the research org than in other OpenAI teams. Agents are handling increasingly complex tasks and succeeding at them more often. The post notes that AI research has many potential bottlenecks, so overall research velocity won’t scale linearly with these specific metrics — but the direction is consistent with the internal perception that agentic tools are materially accelerating progress.

Human researchers still own priority-setting, idea judgment, and the decisions to scale, pause, or deploy systems.

The Next Milestone: Automated AI Researcher by March 2028

OpenAI has set a new target: a full “automated AI researcher” — a system that can originate and execute novel research directions with minimal human direction — by March 2028. That is a qualitatively different claim from an automated intern. An intern runs well-scoped tasks. A researcher decides what to work on.

RSI Is Explicitly on the Table

The post addresses recursive self-improvement (RSI) directly. OpenAI says it does not yet know how to safely reach fully aligned, full RSI — but it frames automated research capability as both a means to build safer AI and a reason for caution. An automated AI researcher can simultaneously be an automated alignment researcher, helping close the gap between capability and safety. Whether to push further “must depend on our ability to preserve human control and on informed democratic choices about the benefits and risks.”

The publication of this post alongside two other OpenAI safety posts on the same day — on agent misalignment monitoring and on Sam Altman’s philosophical essay “An Alien Mind” — appears to be a coordinated transparency package as capabilities cross a milestone the company considers significant.

Benchmark Context

• September 2026 goal: Automated research intern. Status: hit.
• March 2028 target: Automated AI researcher. Status: in progress.
• Current AI research output ratio: 3.1x human staff effort.
• Models driving internal research agents: Codex (GPT-5.5/Astra-class), running in concurrent sessions across researcher workflows.