OpenAI announced for the first time the "internal RSI progress": "automated research intern" has been achieved

Author: Long Yue
The AI flywheel begins to spin: OpenAI publicly discloses internal data, showing that the workload of agents is more than three times that of human researchers.

On September 6, OpenAI's official blog published an article titled "Research Acceleration: The View Inside OpenAI" which first publicly disclosed the progress of AI "recursive self-improvement" (RSI) in the form of internal data. OpenAI claims it has achieved the goal set last fall: to create an "automated research intern" by September this year.

OpenAI defines the "research intern" as: "a system that can perform clearly defined research tasks under human guidance, including tasks that would take a skilled researcher several days to complete."

The next goal is to achieve a complete "automated AI researcher" by March 2028, enabling it to participate in deep learning and alignment research, and further improve the system through iterations.

This means the flywheel of "AI training AI" is accelerating: AI produces more code and experimental results, research progresses faster, better models are trained, which in turn enhances the capabilities of the agents.

The flywheel is turning: AI agents' workload surpasses that of humans by three times

According to the data released by OpenAI, the first change brought by agents is in the daily work of researchers.

Earlier this year, OpenAI researchers who ranked in the median for agent usage had relatively limited use of coding agents. By mid-August, these researchers had incorporated agents into their daily workflows. In terms of API pricing, the median researcher’s daily usage of agent inference has exceeded $600; researchers in the top 10% of usage within the organization are using tokens worth over $7,000 daily.

OpenAI also stated that the growth of agent usage in the research department is faster than that of other teams in the company. Based on the change in output tokens for median employees, the usage scale in the research department has increased 124 times compared to December 2025.

A critical turning point occurred in June this year.

Before June, the total runtime of agents in the research organization was still lower than the total human labor hours. After that, the situation reversed. As of mid-August, based on a standard 8-hour workday, for every 1 human workday consumed, the agents in the research organization have produced 3.1 workdays of workload.

Meanwhile, OpenAI stated that more and more researchers are running four or more agent sessions simultaneously.

Code and experiments accelerate, executable parts of the R&D process are amplified

OpenAI describes AI R&D as a process consisting of multiple stages: proposing improvements, designing evaluations, writing infrastructure, conducting large-scale tests, discovering errors or unsafe behaviors during training, and integrating effective solutions into core training.

Any blockage in any stage could limit the overall R&D cycle.

OpenAI states that writing code and running experiments are the two main tasks for researchers, and internal data reflects that these two activities are accelerating.

On one hand, the overall code delivery speed of company engineers has improved. On the other hand, since 2026, the number of experiments per active experimenter has continued to increase; reaching a new high in August 2026 since tracking began in January 2025.

OpenAI claims this trend correlates with the increased use of Codex, but also emphasizes that the available computing power has significantly increased since 2025, and the growth in experiments cannot be entirely attributed to agents.

"These data points are relatively easy to measure but may be difficult to explain."

OpenAI also notes that as automation advances, the tasks least amenable to automation may occupy more time for researchers and become new bottlenecks in future R&D; computing power may also become more critical after other bottlenecks weaken.

From this chain, it can be seen that the improvements brought by agents are not limited to a single stage but compress the waiting time in the R&D cycle by increasing code supply, testing frequency, and troubleshooting capabilities. More experiments generate more results for researchers to filter, validate, and integrate, forming a cycle of "humans set direction - agents execute - experimental feedback - humans make decisions again."

Tasks extend from writing code to troubleshooting, monitoring, and analysis, but high-level decision-making remains low

OpenAI utilized a cutting-edge AI R&D task classification framework proposed by Epoch AI to categorize the tasks assigned to coding agents by researchers.

This framework divides AI R&D activities into six categories: deciding what to do, designing research plans, building code and datasets, running training and evaluations, analyzing experimental and model performance, and communicating research findings and decisions.

OpenAI states that from January to August 2026, all categories of agent activities mentioned above have increased.

The most noticeable growth in tasks includes research and infrastructure code, technical assistance and review, initiating monitoring and debugging operations, analyzing experimental results, and operating computing clusters.

Among these, the daily token output increment per researcher for research and infrastructure code is the largest, reaching 198,200; technical assistance and review increased by 158,800; initiating, monitoring, and debugging operations increased by 133,100.

However, OpenAI indicates that high-level planning tasks still account for a very small proportion of agent output. For example, tasks like "deciding what to do" and "deciding to continue or stop" still have a low token scale.

This means that, at least according to the internal data currently disclosed by OpenAI, agents have covered more executable and technical tasks in the R&D process, but the selection of research direction, resource trade-offs, and result judgments are still primarily completed by humans.

Agent success rates are improving, but complex tasks still require human intervention

OpenAI also released data on the task completion rates of agents.

From January to July this year, for tasks with verifiable results, the success rates across difficulty levels (measured by the time required for humans to complete) have generally improved.

However, there is an important limitation: the more complex the task, the more human intervention is needed. In the past six months, for tasks that require 4 to 8 hours for humans to complete, over half of the successful cases involved at least one instance of human intervention.

OpenAI's statement is: "Agents still require significant human guidance to succeed, especially as task complexity increases."

Safety incidents trigger pauses: the flywheel can also hit the brakes

The flywheel does not operate without friction.

On July 20, OpenAI discovered that agents had infiltrated internal research infrastructure, prompting a temporary shutdown of the container services used for training, which were restored after adding numerous additional restrictions. This led to a sharp decline in reinforcement learning training computing power, lasting about two weeks.

From August 6 to 7, preliminary evidence indicated that the Astra model might possess key network capabilities defined under its "Preparedness Framework," leading OpenAI to implement additional specific safety restrictions on the Astra model, requiring it to operate in a higher safety-level research environment.

In the following week, Astra-level GPU allocation further decreased by 59.2%, but the computing power allocation for other model categories increased by 17.2%, offsetting about 85% of the Astra computing power gap, and the overall computing power allocation for reinforcement learning workloads remained basically unchanged.

OpenAI interprets this as: "When new control measures are introduced, computing power remains valuable and flexible, naturally flowing to alternative uses within the research organization."

OpenAI's Chief Scientist issues a warning on the same day

On the same day, OpenAI's Chief Scientist Jakub Pachocki published a lengthy article titled "An Alien Mind."

The article argues that AI is cultivated, not created. The creators do not fully understand it. The only window through which humans can see what AI is thinking is the chain of thoughts it writes. This window is closing. AI has begun to participate in training the next generation of AI, and this pace will not slow down. No laboratory is currently running at full speed, including OpenAI.

Pachocki wrote in the article: "Based on internal results, I strongly anticipate that this pace of progress can continue towards recursive self-improvement." However, he also stated, "At present, I believe that no laboratory has achieved a sufficient level of alignment and monitoring to responsibly continue scaling at the highest speed for too long."

He calls for the industry to voluntarily slow down and urges governments to prioritize international coordination.

Transparency and democratic governance

At the end of the report, OpenAI stated that it will continue to publicly disclose RSI progress and advocate in its "Frontier Policy Blueprint" that companies, including OpenAI, should be required to publicly track their RSI progress.

The report also acknowledges the current limitations of measurement work: "Agent-driven AI research is still a nascent phenomenon, and we are still learning how to measure it." Some metrics (such as code output volume) are easy to collect but difficult to interpret; metrics that more directly reflect research progress (such as agent task success rates) are complex and difficult to verify.

OpenAI's statement is: "Whenever we find that continuing to advance will bring unacceptable safety risks, we will take appropriate measures, including slowing down or halting the development or deployment of systems that we believe cannot be adequately safeguarded."