OpenAI AI Agents Complete Nearly Half of Research Tasks Autonomously
OpenAI reports its AI agents now complete 46% of two-to-four-hour research tasks without human intervention, up from 23% in March. The company urges labs to track recursive self-improvement, noting agent usage costs have surged significantly among researchers, highlighting rapid automation in scientific workflows.
Key points
- OpenAI, an artificial intelligence research laboratory, reported that its coding agents autonomously completed 46% of research tasks lasting two to four hours, a significant increase from 23% in March.
- Success rates for longer four-to-eight-hour tasks rose from 13% to 35%, according to the company's second report on AI accelerating its own research published in October.
- Median daily API costs for researchers reached $2,400 in October, a 62% increase from September, while the 90th percentile usage hit approximately $23,000 per day.
- Agent usage time per researcher increased elevenfold to 5.7 agent workdays daily, with more than two-thirds of staff running at least four agents simultaneously at peak times.
- OpenAI stated that the value of agent use in its research organization was about twice the pay bill for researchers in September, though internal use still accounted for only 2% of total research compute.
- The company recommended that AI labs be required to publicly report signs of recursive self-improvement, clarifying that this does not indicate an intent to pursue full autonomous self-improvement.
Acceleration of Internal Research Workflows
OpenAI has released new data indicating a substantial leap in the autonomy of its artificial intelligence coding agents. According to a report titled "Research acceleration: The jagged frontier," published on Wednesday, the company's agents now successfully complete 46% of research tasks that would typically require two to four hours of human effort. This figure represents a dramatic improvement from March, when the success rate stood at 23%. The report marks the second time the organization has documented how AI is speeding up its own internal research processes, following an initial publication in September.
Rising Costs and Usage Metrics
The integration of these agents has led to a sharp increase in operational costs and computational usage within the research division. In October, the median OpenAI researcher utilized approximately $2,400 worth of agent time per day at API list prices. This represents a 62% increase from the $1,500 daily average recorded in September. At the 90th percentile of usage, daily costs reached about $23,000, more than double the $11,000 figure from the previous month. When measured in time rather than currency, agent use per researcher rose elevenfold, climbing from 0.5 to 5.7 "agent workdays" per day. OpenAI defines one agent workday as eight hours of continuous agent running time.
Scaling Concurrent Agent Deployment
The scale of concurrent agent deployment has also expanded rapidly. By mid-October, more than two-thirds of researchers were running at least four agents simultaneously at their daily peak usage. Furthermore, more than 10% of the research staff operated at least 16 agents at once. This surge in parallel processing capability underscores the shifting nature of how researchers interact with AI tools, moving from occasional assistance to continuous, multi-agent collaboration. The value of agent use in OpenAI’s research organization was estimated to be about twice the pay bill for researchers in September, a figure that excludes equity compensation, which the company notes constitutes 70 to 80% of their target pay structure.
Compute Impact and Future Reporting
Despite the high monetary value of API usage, the actual computational load on the company's infrastructure remains relatively modest. Internal use of agents still accounted for only 2% of total research compute. This discrepancy highlights the difference between external API costs and internal model inference resources. OpenAI used this data to advocate for greater transparency in the industry. The company stated that AI labs should be required to track and publicly report signs of recursive self-improvement, a process where AI systems accelerate the development of future AI models. However, OpenAI clarified that this call for reporting does not mean it intends to pursue full autonomous self-improvement.
Key takeaways
- OpenAI agents now complete 46% of two-to-four-hour research tasks autonomously, up from 23% in March.
- Median daily API costs for researchers rose to $2,400 in October, a 62% increase from September.
- Agent usage time per researcher increased elevenfold to 5.7 agent workdays per day by mid-October.
- More than two-thirds of researchers ran at least four agents simultaneously at peak times.
- OpenAI recommends mandatory public reporting of recursive self-improvement signs by AI labs.
- Internal agent use represents only 2% of total research compute despite high API costs.
Sources
The WireByte editorial team synthesises technology news from multiple primary sources, verifies the facts, and links every source. Articles are produced with AI assistance and reviewed under our editorial policy.