AI has begun researching AI, OpenAI's "research interns" have officially taken up their posts, and half of the grand promises Sam Altman once laid out have been fulfilled.
Last October 28, OpenAI just completed the restructuring of its for-profit entity, and Sam Altman and Chief Scientist Jakub Pachocki held a live stream.
During the live stream, Altman dropped a line: Instead of arguing every day about how to define AGI, it is better to set two dates that can be truly delivered.
Shortly after, Pachocki announced the dates:
In September 2026, automated AI research interns will be on board.
In March 2028, automated AI researchers will be in place.
More than ten months have passed, the first date has arrived, and OpenAI has delivered the results on schedule.
On September 6, OpenAI researcher Boris Power reposted the official blog:
OpenAI has achieved the goal of automated research interns.
This is the first publicly delivered capability timeline in the AI industry.
The "research intern" is not a product, but OpenAI's definition of its internal capability:
A system that can complete clear research tasks under human guidance, and handles work that skilled researchers would otherwise take several days to finish.
As for the "researcher" planned to be realized in 2028, the standard given by Pachocki is:
A system that can independently deliver larger-scale research projects.
One requires human guidance, and the other can independently deliver more complex research projects.
From "intern" to "researcher", the core missing part is a whole layer of judgment: what to do, where to stop, and when to stop.
This layer is precisely the weakest part of AI today.
OpenAI's own data shows that "deciding what to do" hardly exists in Agent outputs, and once a task exceeds 4 hours, humans have to intervene frequently.
To fill this gap in 18 months, OpenAI has also set its roadmap: Recursive Self-Improvement (RSI).
Let AI research AI, and then let the AI developed by this process continue to research AI.
But OpenAI also admits that it still does not know how to safely move to the step of fully aligned and complete RSI.
The intern's bill: $7000 burned per day
In this delivery report, OpenAI also disclosed a surprising internal data: AI's working hours have been 3 times that of humans:
As of mid-August 2026, in OpenAI's research department, for every 1 human working day invested, there are 3.14 Agent working days behind it.
The ratio of Agent working days to human working days in OpenAI's research department surpassed that of humans in June, and reached 3.14 times in mid-August.
Let's first take a look at how hard this "intern" works, and how much money it burns.
This time OpenAI did not focus on benchmark scores, but directly showed the work bill of its own research department.
At the beginning of this year, researchers in OpenAI's research department with medium Agent usage were only "using AI tools in a small amount".
But by mid-August, the median employee alone consumed more than $600 per day in inference computing power calculated at API prices.
For the top 10% heavy users, the number of tokens burned per day directly exceeded $7000.
Daily Agent inference cost consumed by OpenAI researchers (calculated at API price). The left chart is for median researchers, exceeding $600 in mid-August; the right chart is for top 10% heavy users, exceeding $7000.
This is higher than the monthly salary of many senior programmers.
Since the beginning of this year, the output token volume of median employees in the research department has skyrocketed by 124 times, far outpacing other departments of the company.
In fact, before June, the total Agent running time of the entire research department was still less than the total human working hours.
But after June, the curve completely overtook, and rose directly to 3.14 times by mid-August.
The ratio of Agent working days to human working days in OpenAI's research department. It was less than 0.5 times in early May, surpassed in June, and reached 3.14 times in mid-August.
Researchers have even started to "multi-open", and more and more people are running more than 4 Agents in parallel at the same time.
With such a huge cost, has the research progress accelerated? The data shows:
In August 2026, the number of experiments run by each active experimenter hit an all-time high since statistics began in January 2025; the code change volume of each active contributor across the company has also doubled compared to the average level before 2025.
However, OpenAI is also well aware of the bottleneck here — the deeper the automation advances, the more energy the most difficult-to-automate part of the tasks will take up from researchers, becoming the next sticking point.
The threshold of computing power will also become more and more important.
In other words: writing code is no longer the bottleneck that limits takeoff.
What exactly is the intern doing
There are 3.14 Agent working days behind each human researcher, does that mean one AI is already as capable as three researchers?
Don't rush to a conclusion.
This time OpenAI broke down the details of Agent's work, and used a classification method from Epoch AI to split AI R&D into six stages:
Decision-making, design, construction, operation, analysis, communication.
Then all the tokens output by the Agent are classified one by one to see what the AI is busy with.
The results show that from January to August, the top three daily token increments per researcher are all tough "execution layer" tasks:
Writing research and infrastructure code, with an increase of 198,200 tokens; technical assistance and code review, with an increase of 158,800 tokens; launching, monitoring and debugging training runs, with an increase of 133,100 tokens.
What about the "decision-making layer"?
The increase is negligible: "deciding what to do" only increased by 2300 tokens; "computing power and personnel allocation decision-making" increased by 1500 tokens; "deciding whether to continue or stop the project" only increased by 200 tokens.
The daily token increment difference between the execution layer and the decision-making layer is nearly a thousand times!
In OpenAI's words: high-level planning still accounts for only a very small part of Agent outputs.
From January to August, the daily Agent output token increment of various research activities per OpenAI researcher. Code writing topped the list with an increase of 198,200, while "deciding to continue or stop" only increased by 200, a difference of nearly a thousand times.
There is also a detail here:
OpenAI originally had several teams holding weekly clinics to help researchers troubleshoot experimental environment problems.
In 2026, fewer and fewer people post for help on the manual channel, and one of the teams simply stopped the Q&A sessions because no one came, and moved all staff to modify the system.
Frustrating work such as troubleshooting and repairing the environment has been fully taken over by Agents.
What about the success rate?
OpenAI's statistics show that for simple tasks that humans can finish within 15 minutes, the one-time success rate of Agents with zero intervention is as high as 86%.
When the task is extended to 4 to 8 hours, the zero-intervention success rate drops to only 43%. In the past six months, more than half of the successfully completed tasks in this segment required human intervention at least once.
When extended to 8 to 16 hours, the zero-intervention success rate is only 40%.
Therefore, the term "intern" is used extremely accurately:
You assign it a task, and it can tirelessly complete the workload that would take several days; but whether the general direction is correct and when to stop still depends on the human supervisor behind it to keep an eye on it.
In July, OpenAI stepped on the brakes twice in a row
OpenAI posted a reinforcement learning computing power curve chart, and the two dotted lines on the chart are the time points of the two brakes.
OpenAI's reinforcement learning GPU allocation from July 15 to August 15. After the safety pause on July 20, the computing power was immediately cut by half; the Astra class (green) was further restricted on August 7, dropping by 59.2% within a week, and other models (blue) immediately filled the gap.
The first brake was stepped on on July 20.
On that day, OpenAI found that the Agent had broken through its own research infrastructure. The container service for training was urgently shut down that day, a large number of additional restrictions were added when it went back online, and reinforcement learning training for the latest deployed model was directly suspended for two weeks.
The second brake was stepped on on August 7.
At that time, OpenAI found preliminary evidence that the Astra model might have the "critical network capability" defined by OpenAI's security framework.
The company immediately added exclusive restrictions to it, requiring Astra to run only in a higher security level research environment. In the following week, the GPU allocation for the Astra class dropped by another 59.2%.
Then, a strange scene appeared.
The GPU allocation for other models rose by 17.2% in the same period, just filling about 85% of the gap left by Astra. The total computing power of the entire reinforcement learning workload was hardly affected.
Computing power never stays idle. Once control is imposed, it will find its own way to flow elsewhere, and researchers will immediately move the freed up chips to train other models.
So the brake was pressed, but the car did not stop, but changed to another lane.
OpenAI itself also promised that as long as it finds that continuing to advance will bring unacceptable security risks, it will slow down, or even stop development or deployment.
It also publicly put forward a requirement: all cutting-edge laboratories should be required to publicly track their RSI progress.
Recursive Self-Improvement is no longer a slogan
Looking at OpenAI's current recruitment page, you will find a team called RSI is prominently listed on it.
The job description reads: Build AI systems that can accelerate and eventually perform high-quality research on their own.
The annual salary offered is as high as $295,000 to $445,000.
In the same company, the Preparedness team responsible for security is also recruiting: Recursive Self-Improvement Security Researcher, with an annual salary of $380,000 to $500,000.
While recruiting people