Altman’s Timeline from a Year Ago Has Come True Right on Schedule.
On September 6, OpenAI published a major blog post titled “Accelerating Research: An Inside Look at OpenAI,” publicly revealing internal data on AI building AI for the first time.
The post opened by highlighting a key milestone: before September this year, OpenAI had successfully deployed an automated AI research intern. The next target is to build a fully automated AI researcher by March 2028.
Around the same time, Nvidia CEO Jensen Huang posted on X congratulating OpenAI on the launch of GPT-6 Astra and declaring that AGI has arrived. From ChatGPT to o1 and now Astra, he noted, it took just four years. Huang also revealed that Astra was trained using more than 100,000 NVIDIA Grace Blackwell NVLink72 GPUs, with another 400,000 GPUs set to be delivered to OpenAI.
That same day, OpenAI Chief Scientist Jakub Pachocki published his essay “An Alien Mind,” acknowledging that machine intelligence has surpassed human capabilities and warning that humanity is creating an increasingly incomprehensible “alien mind” that requires an urgent global slowdown.
On one side, evidence that recursive self-improvement is accelerating. On the other, an alarm calling for the brakes. Two completely opposing messages—both coming from within OpenAI.
OpenAI has disclosed a set of core internal figures that it has rarely shared before.
Based on API pricing, the median researcher spends more than $600 per day on agent inference, while the top 10% of researchers spend more than $7,000 per day. Agent usage in the research division is growing far faster than in other parts of the company, with median employee token output reaching 124 times its level at the beginning of the year.
OpenAI converted the total runtime of all research agents into standard eight-hour workdays and compared it with human workdays.
Before June 2026, agents had yet to surpass humans in total working hours. By mid-August, however, the ratio had flipped to 3.1 to 1.
For every human researcher sitting at a desk, 3.1 AI workdays are now running in parallel behind them.
Parallelization has been the biggest driver. At the beginning of the year, roughly 30% of researchers were running four or more agents simultaneously. By mid-August, that figure had surged to more than 70%.
OpenAI breaks the research process into six stages: decision-making, design, building, execution, analysis, and communication.
From January through August 2026, agent token output surged across all six categories. But most of the growth was concentrated in the middle of the workflow—writing research and infrastructure code, providing technical assistance and reviews, and launching and monitoring training jobs.
At the strategic level, where decisions are made about resource allocation and major directions are set, agents still have very little influence.
Another telling sign is the decline of internal help channels.
A once-busy research help channel that received around 20 new posts a day now sees only single-digit activity. The problems haven’t disappeared. The people asking for help have simply shifted from humans to agents.
Complexity presents another constraint.
For medium-sized tasks lasting four to eight hours, more than half of successful runs require at least one human intervention.
OpenAI’s own assessment is blunt: agents still require substantial human guidance to succeed, particularly as tasks become more complex.
OpenAI makes a counterintuitive observation in the report: as automation advances, the tasks that are hardest to automate will consume an increasingly large share of researchers’ time—and eventually become a new bottleneck for AI development.
The report points to the same limiting factor: compute.
OpenAI team member Kevin Liu said on X that recursive self-improvement, or RSI, is likely to become a key driver of AI capability gains over the next several years, but that it currently exists only within a small number of leading AI labs.
Back in October, Altman laid out a timeline during a livestream: build an automated AI research intern by September 2026, followed by a fully automated AI researcher by March 2028.
The first milestone has now been reached.
OpenAI says it is making strong progress toward the March 2028 target.
And whether or not the industry agrees with Huang’s declaration that “AGI has arrived,” it points to at least one undeniable reality:
The flywheel of AI building AI is already spinning.
The next question is who gets to control its speed.
