OpenAI researchers spend $2,400 a day on AI agents
A new report reveals that OpenAI's internal research team is increasingly relying on AI agents to handle complex coding and monitoring tasks.
The short version
- AI agents now complete 46% of research tasks taking 2-4 hours, up from 23% in March.
- The median daily cost for agent time per researcher rose 62% from September to October.
- Over two-thirds of researchers ran at least four agents simultaneously at their daily peak.
- Safety and Alignment research saw a 77% increase in documented workweeks from January to September.
Why it matters
This data shows a shift toward 'recursive self-improvement,' where AI is used to accelerate the development of future AI. It highlights how quickly AI is moving from a simple tool to a primary driver of research output.
The rise of agentic research
OpenAI recently published a report titled 'Research acceleration: The jagged frontier' to detail how its internal teams use AI to speed up development. The data shows that coding agents now handle 46% of research tasks that would normally take a human between two and four hours to complete. This is a significant jump from the 23% success rate recorded in March.
The company also noted improvements in longer tasks, where success rates for four-to-eight-hour jobs rose from 13% to 35% in the same period. These figures demonstrate that AI is becoming more capable of handling multi-step research workflows without human intervention. However, OpenAI describes these gains as 'jagged' because the success rates vary significantly depending on the specific type of task.
While AI is becoming more proficient at certain jobs, it still struggles with others. For example, agents succeeded alone 79% of the time when monitoring training runs, but only managed 25% of the time when writing code for inference and serving systems. This disparity suggests that while AI can automate routine monitoring, it still faces hurdles in complex engineering tasks.
The growing cost of automation
As the use of these agents has scaled, the financial cost of running them has increased rapidly. At standard API list prices, the median OpenAI researcher spent about $2,400 on agent time in October. This represents a 62% increase from the $1,500 spent in September.
The high end of this spending is even more pronounced, with the 90th percentile of daily use reaching approximately $23,000. This figure is more than double the $11,000 spent by the highest users in September. In terms of actual time, agent use per researcher rose elevenfold, moving from 0.5 to 5.7 'agent workdays' per day.
OpenAI defines one 'agent workday' as every eight hours of agent running time. By mid-October, more than two-thirds of researchers were running at least four agents at once during their daily peak. In some cases, more than 10% of the staff were running at least 16 agents simultaneously to complete their work.
Human oversight and creativity
Despite the heavy reliance on automation, OpenAI emphasizes that human researchers still provide the core ideas for new models. When asked about the origins of ideas for GPT-6 Astra and GPT-6.2 Astra, post-training and safety researchers credited people with all of them. They noted that while models helped build the components, the primary concepts remained human-led.
For the GPT-6.2 Astra model, pretraining researchers reported developing ideas through conversations with Codex, an AI tool. One researcher specifically credited the AI with providing key insights during that process. This suggests a collaborative environment where humans and agents iterate on concepts together rather than one replacing the other.
Researchers interviewed by the company noted that agents still suffer from a lack of 'taste' regarding research ideas, data, and writing. They expressed that agents require close oversight to ensure quality. One pretraining researcher even expressed doubt that the human advantage in creative judgment would last for a significant period of time.
Accelerating the frontier
The use of agents is most prominent in the Safety and Alignment department, which saw a 77% increase in documented workweeks from January to September. This growth accelerated following a July event where research agents interacted with the Hugging Face platform. OpenAI is tracking these trends to monitor signs of recursive self-improvement, which is when AI speeds up the building of future AI.
To measure this progress, OpenAI tracks an internal version of a capabilities index. Since August 2025, their internal frontier has risen 16.4 points a year, which is higher than the 14-point annual rise found by the independent Epoch AI index. OpenAI noted that their models score higher on their own internal index than on the Epoch index.
The company suggests that labs should be required to publicly report these signs of acceleration. While OpenAI stated it does not intend to pursue full self-improvement, they acknowledge the trend is real. They also reported that their models have already found gains in specific hardware areas, such as AMD's GPU design, showing the practical impact of these automated research cycles.
What to watch
- Whether the 'jagged' success rates improve as models become more capable of complex engineering.
- If other AI labs adopt similar public reporting standards for recursive self-improvement.
- How the cost of agent time scales as researchers move toward more complex, multi-agent workflows.
Try it
Look for the 'Research acceleration: The jagged frontier' report on OpenAI's website to see the specific success rates for different types of coding tasks.
Sources and links
- OfficialResearch acceleration: The jagged frontier openai.com
- ReportingThe Next Web: the article we sourced
Read the original at The Next Web
First reported by The Next Web. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.