OpenAI Researchers Spend $2,400 Daily as AI Agents Complete 46% of Tasks
OpenAI researchers now spend an average of $2,400 daily on AI agents, with top users exceeding $23,000. Recent data shows agents successfully complete 46% of two-to-four-hour research tasks without human intervention, a significant rise from 23% in March. While experiment volume increased by 68%, capability gains remain modest, and performance varies drastically across tasks. Notably, the Codex model assisted in identifying key insights for GPT-6.2 Astra, marking a potential foothold for AI in research taste.
Editorial responsibility
- No named human review is recorded for this page.
- Reports are grouped by semantic similarity and deterministic rules. Language models may assist titles, summaries, translation and cross-source analysis; the page reads the event directly, while its address stays stable when the title changes.
- Summary covers the current reports
Cross-source coverage
Reporting timeline
OpenAI Researchers Burn $2,400 Daily on AI Agents as GPT-6.2 Astra Emerges
A recent report indicates that typical OpenAI researchers now spend approximately $2,400 daily on AI agents at API prices, with the top 10% exceeding $23,000 per day. While these agents run for durations equivalent to 5.7 full-time employees, a significant disparity exists between cost and output; experiments increased by 68%, yet capability gains only slightly outpaced industry trends. The analysis highlights that agents succeed in 79% of monitoring tasks but fail in 75% of writing inference and serving code, suggesting these weaker areas may limit future research speedups. Regarding model development, key ideas for GPT-6 Astra originated from human researchers, whereas for GPT-6.2 Astra, pretraining researchers credited Codex with identifying key insights, marking what is described as AI's first reported foothold in research taste.
Read sourceOpenAI looked at how its own researchers use AI to build better AI. They tracked agent usage, success on delegated tasks, experiment volume, and model benchmark scores. The frontier is “jagged” because agents can do
OpenAI analyzed how its own researchers utilize AI agents to develop better artificial intelligence systems. The study tracked key metrics including agent usage frequency, success rates on delegated tasks, experiment volume, and model benchmark scores. A significant finding revealed that 46% of delegated tasks lasting two to four hours were completed successfully without human intervention, a substantial increase from 23% in March. For longer tasks ranging from four to eight hours, the success rate rose from 13% to 35%. The report characterizes the current frontier of AI capabilities as 'jagged,' noting that while agents can autonomously handle specific parts of complex jobs, they still struggle significantly with other components. Analysts caution that these results reflect changes in task selection by researchers rather than a clean model-to-model benchmark comparison.
Read sourceOpenAI Agents Complete 46% of Research Tasks Without Human Help
New data from OpenAI indicates a significant leap in AI agent capabilities, with agents now successfully completing 46% of research tasks that typically require two to four hours of human effort, achieving this without any human assistance. This represents a substantial increase from March, when the completion rate was only 23%. The report highlights heavy integration of these tools by researchers, noting that the median OpenAI researcher utilizes $2,400 worth of agent services daily at API prices, while the top 10% spend up to $23,000 per day. Additionally, specific models like GPT-6.2 Astra and Codex are credited with identifying core insights during collaborative workflows with pretraining researchers. These figures suggest rapid progress toward artificial general intelligence milestones, though the source text frames this as an observation on the proximity to Recursive Self-Improvement (RSI) rather than a formal announcement.