OpenAI has published internal data showing how deeply autonomous AI agents have already penetrated its own research operations. According to the company's figures, agents now perform 3.1 times the workload of human researchers – a ratio that has grown steadily since June. The company declares its self-imposed goal of an "automated research intern" achieved. Yet Chief Scientist Jakub Pachocki delivers an unusually stark warning: no lab has adequately solved the control problem for such systems.
Key Facts
- 3.1 agent work-days per human work-day (as of mid-August)
- Token output of the median researcher has increased 124-fold since December 2025
- The median researcher consumes daily inference worth over $600 at API prices, the 90th percentile over $7,000
- OpenAI plans a fully automated AI researcher by March 2028 – but humans retain decision-making authority over priorities, results, and scaling
Measurements Without External Audit
OpenAI released two coordinated texts: a blog post with internal metrics and Pachocki's essay "An Alien Mind" – both three days after the unveiling of GPT-6 Astra. The core message: the company is making rapid progress on recursive self-improvement (RSI) and considers that dangerous.
However: all data comes from within the company itself. OpenAI mentions no independent verification. The achievement of the "automated research intern" is justified merely with "according to our measurements" – detailed validation remains absent.
Where Agents Dominate – and Where They Don't
The nature of delegated work is shifting noticeably. Classified according to Epoch AI's taxonomy, all categories of research work are growing, most strongly:
| Task Type | Trend |
|---|---|
| Writing research and infrastructure code | Strong growth |
| Technical support | Strong growth |
| Monitoring training runs | Strong growth |
| High-level planning decisions | Minimal share |
Strategic decisions remain with humans – for now. OpenAI itself cautiously frames the metrics: they are "relatively easy to collect but hard to interpret, because their relationship to research progress is uncertain." Overall progress likely grows slower than individual metrics, because the least automatable tasks become the bottleneck.
The Chief Scientist's Warning
Pachocki makes clear in his essay: the speed of automation is a problem, not just an opportunity. The implicit message: OpenAI is scaling faster than its safety infrastructure allows. Humans still set research priorities, evaluate results, and decide on pauses – but as systems become more autonomous, the question of whether human control still holds becomes critical.
What This Means for German Companies
These figures are a wake-up call for decision-makers in Germany. They show: productivity gains from autonomous AI are no longer a future vision but present reality – at least for tech leaders. Companies deploying AI agents should ask whether their control mechanisms and alignment processes keep pace with this speed. At the same time, Pachocki's warning is a signal: scaling too fast without understanding risks risks losing control quickly. For German research institutions and companies, this means: now is the moment to build governance structures for autonomous AI systems – before they become routine.
Sources
Editorially owned by Ideal Syka. Sources and method: Newsroom & method. Tips and corrections: ai@i6eal.de.
