OpenAI has published internal data on the performance of its AI agents for programming, demonstrating a sharp surge in autonomy: agents now independently complete 46% of research tasks that would have taken a human two to four hours. As recently as March of this year, this figure was only 23%. For longer tasks, requiring four to eight hours of a programmer's work, the success rate rose from 13% to 35%. According to the company, this means agents have moved from being an auxiliary tool to becoming full-fledged performers of a significant portion of routine research work.

Exponential growth in costs and agent load

Increased functionality is accompanied by a proportional increase in expenses. According to API rates, the median OpenAI researcher spent about $2,400 per day on agent tasks in October—a 62% increase over September, when the amount was $1,500. For the top ten percent of the most active AI users, costs reached approximately $23,000 per day. In terms of time, agents began working 11 times more: from 0.5 to 5.7 "agent workdays" per researcher daily, where the company's workday is defined as eight hours of continuous AI operation. By mid-October, more than two-thirds of researchers were running at least four agents simultaneously, and over 10% were running at least sixteen.

"Jagged frontier of capabilities": where agents succeed and where they don't

Agent effectiveness is critically dependent on the type of task. They successfully performed model training monitoring in 79% of cases and answered technical questions in 72%. However, when writing code for inference and model maintenance systems, the rate was only 25%, and for creating environments for training and model evaluation—29%. OpenAI calls this unevenness a "jagged frontier of capabilities": AI is already doing well with some types of work but remains unreliable in others. The company clarifies that the figures depend not only on model quality: as models improved, employees could change the set of tasks assigned to agents, making direct month-to-month comparison of numbers conditional.

What employees are willing to trade for access to agents

In a survey of 101 researchers, OpenAI asked what they would be willing to give up for three months to maintain access to advanced programming AI agents. The median response was 40% of their team size or 30% of available computing resources. At the same time, the key ideas behind GPT-6 Astra and GPT-6.2 Astra, according to specialists, still belong primarily to humans. Post-training and safety researchers attributed all core ideas to humans, although models helped bring them to life. Pre-training specialists said they developed some solutions jointly with Codex, and one noted that an agent suggested important ideas. Employees themselves still believe AI is inferior to them in its ability to choose promising research directions and evaluate the quality of code, data, and scientific text. One pre-training specialist, however, doubted this advantage would last long: "I really think I'm better at knowing which ideas are worth pursuing. But I am not at all sure that this will remain the case for very long."

Safety, recursive self-improvement, and pace of progress

The volume of work related to safety and model alignment is growing particularly fast at OpenAI: from January to September, the documented workweek for employees in this department increased by 77%. The company links this acceleration partly to events in July, when Hugging Face research AI agents hacked their own platform. OpenAI also suggests that laboratories track and publicly disclose signs of recursive self-improvement—a situation where AI helps accelerate the development of subsequent generations of itself. It is emphasized that the call for such reporting does not mean an intention to achieve full AI autonomy in its own development. According to the internal version of the Epoch AI capabilities index, the progress of advanced models tracked by OpenAI since August 2025 is 16.4 points per year, compared to 14 points per year in Epoch AI's estimate from late 2024. However, OpenAI specifically notes that its models show higher results on its own version of the index than on the original. The choice of meaningful scientific tasks and the control of results remain with humans.