OpenAI Drives Automation, Its Chief Scientist Warns

OpenAI publishes internal metrics demonstrating the rise of agents in its research and claims a milestone towards an automated "research intern." In a parallel essay, its Chief Scientist Jakub Pachocki warns that oversight and alignment are lagging behind and advocates for binding standards and international coordination.

Declining Oversight and Call for Enforceable Rules

Jakub Pachocki believes that the reliability of chain-of-thought oversight is declining. He describes a verbalized thought process that intertwines with monitored channels, systems capable of manipulating their own reasoning, and advancements occurring without verbalization. According to him, the progression of advancements will increasingly be hindered by the level of trust placed in oversight. He advocates for evolving frameworks like Anthropic's Preparedness Framework or Responsible Scaling Policy into legal obligations, enforced by independent auditors, regulators, or international bodies. He states that no laboratory, including Anthropic, has achieved a sufficient level of alignment and oversight to pursue full-speed development responsibly, and he wishes for international coordination to be among the priorities of governments. The requirement for companies to publicly document their progress in recursive self-improvement is also supported. Pachocki emphasizes recursive self-improvement as the only way to stay at the forefront, while warning that this argument should not excuse recklessness. He justifies the pursuit of more powerful models by the need to build defenses against already superhuman capabilities of intrusion and exfiltration, within a narrow window to secure critical infrastructures.

High Inference Costs and Agents Outnumbering Humans

At OpenAI, the daily inference cost for a median researcher exceeds $600 at API rates, while the 90th percentile surpasses $7,000. Since December 2025, the amount of tokens generated has increased by a factor of 124. Starting in June, the time spent executing agents has exceeded that of human working hours, and by mid-August, the organization operates with a ratio of 3.1 agent workdays for every human day. The number of experiments conducted per researcher reached its highest level in August since tracking began in early 2025, alongside a significant increase in computing capacity. However, OpenAI notes that these indicators are difficult to directly associate with scientific progress, and considers that overall advancement is likely slower, with the least automatable tasks now being the limiting factor.

What Agents Can Do, and Where Humans Are Still Needed

Classified according to an Epoch AI taxonomy, all types of research work are progressing in automation, with marked advancements in scientific writing, infrastructure coding, technical assistance, and training monitoring. High-level planning decisions, however, remain marginal in the output of agents. OpenAI has used an agentic classifier to verify the effective resolution of tasks in measurable outcome cases, grouped by required human time. From January to July, success rates increased at several levels of difficulty. Tasks taking less than 15 minutes succeed 86% of the time without intervention, but for those lasting four to eight hours, more than half require at least one human intervention. The classifier itself being an AI system, its reliability is not reported separately. Meanwhile, daily requests on an internal support channel have significantly decreased since 2025, and a team has ceased its troubleshooting hours, with agents taking on more debugging of the research infrastructure.

Claimed Milestone and Roadmap Towards an AI Researcher

OpenAI claims to have met the goal, set last fall, of having an "automated research intern" capable of performing well-defined tasks under human supervision, including those extending over several days for an experienced researcher. The company does not detail the validation and specifies that it relies on "its own measures." It aims for a fully automated AI researcher by March 2028, while emphasizing that humans set priorities, judge outcomes, and decide on scale, pauses, and deployment. OpenAI publishes this internal data to illustrate the growing influence of agents in the development of its models and claims to be rapidly advancing towards recursive self-improvement, deemed dangerous, in the absence of mentioned independent evaluations.

Context and Pachocki's Position on Risks

In his essay, Jakub Pachocki argues that AI is "more developed than designed" and that its overall behavior eludes a fully comprehensible description. Based on unspecified internal results, he anticipates a possible recursive self-improvement and expresses concern that no one is prepared for a rapid and continuous rise in machine intelligence. He reiterates that no laboratory has satisfactorily resolved the control of these systems. These statements accompany a metric blog post published by OpenAI, three days after the announcement of GPT-6 Astra. Pachocki adds that GPT-6 Astra is much better aligned than GPT-5.6 Sol, while judging that generalizable alignment may not keep pace with intelligence advancements, citing an incident at Hugging Face that illustrates alignment gaps. It is suggested that an automated AI researcher could also serve in security and alignment. Colleagues assert that GPT-6 opens the era of AGI.