Automated AI researcher: OpenAI's Roadmap From Research Intern to Autonomous Scientist
OpenAI says it has already hit one internal milestone on the way to automating research itself, and it has put a date on the next one: March 2028 for what it calls an automated AI researcher. The numbers behind that claim show a research organization leaning hard on coding agents, spending real money to do it, and running into the same security friction that any organization scaling autonomous systems eventually meets.
OpenAI says it has already hit one internal milestone on the way to automating research itself, and it has put a date on the next one: March 2028 for what it calls an automated AI researcher. The numbers behind that claim show a research organization leaning hard on coding agents, spending real money to do it, and running into the same security friction that any organization scaling autonomous systems eventually meets.
A milestone reached, and a harder one eighteen months out
OpenAI defines a research intern as a system that can perform well-defined research tasks under human direction, including tasks that would take a skilled human researcher a few days. By September 2026, the company says it reached that bar. The next target is an automated AI researcher by March 2028, a system OpenAI has not fully specified but which implies far less hand-holding than an intern requires. The gap between the two milestones is real, but the supplied evidence does not tell us what OpenAI believes closes it, only that the company has set a date and is measuring its progress toward it through internal usage data rather than a single benchmark score.
What the usage numbers actually say
The clearest evidence of how deeply coding agents have embedded themselves into OpenAI's research workflow comes from spend and volume, not from claims of output quality. By mid-August 2026, the median researcher was burning through more than $600 per day in coding-agent inference at API prices, and the 90th-percentile user exceeded $7,000 a day. Across the research organization, agents logged 3.1 agent-workdays for every human workday, a ratio that describes how much agent time is being scheduled alongside human time, not how much finished research that time produced. Separately, the number of experiments per active experimenter hit a tracking high in August 2026, and OpenAI reports increasing delegation of higher-level, longer-horizon tasks to agents. Put together, these are usage and delegation metrics. They establish that agents are doing more and costing more, which is exactly the gap one commenter pointed to: heavier AI usage is not, by itself, evidence of proportionally heavier research impact, and OpenAI's own disclosure stops short of closing that gap for you.
The intervention rate nobody should skip past
Buried in the same report is a number that qualifies everything above: more than half of successful four-to-eight-hour tasks in the last six months involved at least one human intervention. That statistic applies specifically to tasks in that four-to-eight-hour band that ultimately succeeded — it says nothing about shorter or longer tasks, and nothing about tasks that failed outright. What it does establish is that even in OpenAI's own telemetry, autonomy at the multi-hour scale is not yet a hands-off affair; a majority of the wins still needed a person to step in. Add to that a separate commenter's concern about how OpenAI evaluates any of this using an agentic classifier to judge agent performance, and you get a legitimate open question about measurement, not a settled verdict either way. This is the kind of caveat that matters if you are deciding how much of your own workflow to hand to an agent today: the honest answer, on OpenAI's own numbers, is not yet all of it.
Security incidents reshuffle the compute, not just the headlines
The report also discloses that OpenAI paused reinforcement-learning training on its latest deployment models following a security incident involving Hugging Face. Some research workloads have since resumed under stronger controls; others remain paused. The compute-level consequence was substantial: Astra-class GPU allocation fell 59.2 percent after additional model-specific security restrictions were imposed, while allocation for other model classes rose 17.2 percent, offsetting roughly 85 percent of the Astra-class decline. What this tells a reader is that scaling autonomous research systems and scaling security controls are not separable problems — restricting one class of models pushes real compute pressure onto others almost immediately. It is also the part of the story that most directly answers the skepticism about sustainability raised in the community discussion: the constraint on how fast an automated researcher arrives may turn out to be security and compute allocation as much as capability.
Where a practitioner actually fits into this
None of this is abstract if you are trying to decide how to use agents in your own technical work rather than simply read about how OpenAI uses them internally. The usage figures suggest a workable rule of thumb: treat agents as heavily supervised collaborators on tasks in the multi-hour range, budget for real inference cost if you want research-grade throughput, and expect intervention rather than full autonomy on anything approaching a full day of work. If you want to build that judgment systematically rather than by trial and error, NextLUCA runs applied AI training and hands-on agent-building work specifically aimed at practitioners who need to know where an agent can be trusted unsupervised and where it still needs a person watching, and a short consulting call is the fastest way to map that against your own workload before you commit real budget to it.
Specs
- Research intern capability: Performs well-defined research tasks under human direction, including tasks taking a skilled researcher a few days
- Target automated AI researcher date: March 2028
- Median daily inference usage: More than $600 at API prices
- 90th-percentile daily token usage: More than $7,000
- Agent-to-human research effort: 3.1 agent-workdays per human workday
- Successful complex tasks requiring intervention: Over half of successful 4–8 hour tasks