OpenAI said on September 6, 2026 that it has reached the goal it announced last fall: fielding an "automated research intern" by September of this year. In a new blog post titled "Research acceleration: The view inside OpenAI," the company also disclosed for the first time just how deeply AI agents have been woven into its own research organization — by mid-August, the org was using 3.1 agent-workdays of effort for every single human workday.
The milestone, first reported by Engadget and Unite.AI, is the clearest quantitative picture yet of what OpenAI means when it talks about AI accelerating AI research. For readers following the broader story, this fits into our continuing AI research coverage of how frontier labs are restructuring work around agents.
What 'Automated Research Intern' Actually Means
The blog post defines the target precisely: a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days to complete. According to OpenAI's own measurements, that bar has now been met, roughly on the schedule the company set when it announced the goal last fall.
The next target is considerably more ambitious. OpenAI says it is making "strong progress" toward an automated AI researcher by March 2028 — a system that works under human supervision to push forward deep learning and alignment research, enabling iterative self-improvement. The company is careful to note that humans still set research priorities, judge which ideas and results are worth pursuing, and decide whether to scale, pause, or deploy any given system.
The Numbers Behind the Claim
The most striking figures in the post describe how radically the daily work of OpenAI's researchers has changed over the course of 2026:
- 3.1 agent-workdays per human workday. Before June 2026, total agent runtime across the research organization was still below total human labor. By mid-August, agents were contributing more than three times the effort of the humans directing them, measured against a standard eight-hour day.
- $600+ per day for the median researcher. At the start of the year, the median researcher ranked by agent usage was only experimenting with coding agents. By mid-August, the median researcher was running agents daily and consuming more than $600 per day of inference at API prices.
- $7,000+ per day at the 90th percentile. The heaviest users in the research organization now burn through more than $7,000 of tokens per day, running highly concurrent workflows with four or more agents working simultaneously.
- Record experiment volume. The number of experiments per active experimenter hit an all-time high in August 2026, since tracking began in January 2025.
OpenAI attributes the productivity jump primarily to the adoption of Codex, its coding agent, while acknowledging that available compute has also grown substantially since 2025 — so the gains are a combination of more agents, better agents, and more hardware to run them on.
What the Agents Are Actually Doing
To classify agent activity, OpenAI used a research-work taxonomy published by Epoch AI, inspired by the O*NET occupational classification system. It breaks AI research and development into six phases: Decide, Design, Build, Run, Analyze, and Communicate.
All six categories of agent activity increased between January and August 2026. In January, the dominant use was research and infrastructure code; that category has kept expanding, with notable growth in technical help and in monitoring experiment runs. High-level planning, by contrast, remains a minimal fraction of agent output tokens — the strategic decisions still sit with humans.
Anecdotally, the post reports, OpenAI's researchers have found coding agents excel at troubleshooting internal research infrastructure, and multiple teams that previously held office hours for infrastructure issues no longer need to.
Why the Milestone Matters Beyond OpenAI
The announcement lands at a sensitive moment. Lawmakers, safety researchers, and competitors are all watching for evidence that AI systems can meaningfully accelerate AI research itself — the mechanism most likely to compress timelines toward more capable systems. OpenAI's own chief scientist published a companion essay the same day warning that no lab has yet solved alignment well enough to keep scaling at full speed, a striking juxtaposition with the productivity numbers.
The economic implications are just as notable. If a single research organization is consuming thousands of dollars per researcher per day in inference — and treating it as a bargain — that signals where enterprise AI spending is heading. Agent-workdays are becoming a purchasable input, like compute itself.
Skeptics will point out that OpenAI graded its own homework: the "automated research intern" standard is defined and measured internally, and the 3.1:1 ratio counts both directly launched agents and downstream subagents. The company's figures also blend coding, monitoring, and analysis work of very different difficulty levels.
Still, the direction is unambiguous. Eight months ago, agent time lagged human time inside one of the world's best-funded AI labs. Today it outweighs it three to one — and the lab's leadership is already framing the next milestone, the automated AI researcher by March 2028, as a matter of "when," not "if."
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
