According to findings published Wednesday in Research acceleration: The jagged frontier, OpenAI's coding agents have substantially improved their ability to handle research work without human intervention. The report marks the company's second public assessment of how artificial intelligence accelerates its own research operations, following an initial study released in September.

OpenAI contends that research institutions ought to establish mandatory systems for tracking and disclosing evidence of recursive self-improvement, the phenomenon where AI systems enhance the development of subsequent AI systems. The company clarifies, however, that advocating for such transparency does not signal its own commitment to pursuing full self-improvement capabilities.

$2,400 a day per researcher

During October, the typical OpenAI researcher consumed approximately $2,400 worth of agent services at standard API pricing. This represents a 62 percent increase from the $1,500 figure recorded in September. Among the top 10 percent of users, daily expenditure climbed to roughly $23,000, nearly triple the $11,000 observed the previous month.

When calculated by duration rather than cost, agent utilization per researcher expanded dramatically, growing from 0.5 to 5.7 "agent workdays" daily, with each workday representing eight hours of continuous agent operation. By mid-October, over two-thirds of the research staff were simultaneously running a minimum of four agents during peak usage periods, while more than one-tenth operated at least 16 agents concurrently.

In September, the aggregate value of agent services deployed across OpenAI's research division roughly equaled twice the organization's researcher payroll, though this calculation excludes equity compensation, which OpenAI indicates comprises 70 to 80 percent of researcher target compensation. Despite this expansion, internal research compute allocation to agent use remained modest at just 2 percent.

Strong at monitoring, weak at serving code

Performance metrics for OpenAI's research agents demonstrate considerable variation across different task categories, prompting the company to characterize its progress as uneven. Agents functioned independently with 79 percent success when overseeing training operations and 72 percent when responding to technical inquiries. However, they achieved only 25 percent autonomous completion when writing code for inference and serving infrastructure, and 29 percent when constructing evaluation and training systems.

OpenAI acknowledges an important limitation in interpreting these results. As agent capabilities advanced, researchers may have selectively reassigned different categories of work to the systems, meaning the success metrics reflect both genuine performance improvements and shifts in task assignment patterns.

In a questionnaire administered to 101 research personnel, OpenAI inquired what each respondent would forgo to retain access to advanced coding agents for a three-month period. The median response indicated a willingness to sacrifice 40 percent of team staffing or 30 percent of computational resources. OpenAI cautions that such self-reported data may contain inaccuracies and that the sampling methodology was informal rather than systematic.

Ideas still come mostly from people

OpenAI surveyed leading researchers regarding the origins of core concepts underlying GPT-6 Astra and GPT-6.2 Astra. Post-training and safety researchers attributed all major ideas to human contributors, with AI systems playing a supporting role in implementation. For GPT-6.2 Astra, pretraining researchers reported developing ideas through collaborative interaction with Codex, with one researcher acknowledging the system's contribution to critical breakthroughs.

During interviews, researchers emphasized that agents currently demonstrate poor judgment regarding research directions, code quality, data selection, and written output, requiring substantial human supervision. One pretraining researcher expressed skepticism about whether humans would maintain this evaluative advantage.

I do think I have more taste in terms of the ideas. But I have no confidence that will continue to be true for any significant period of time.

OpenAI pretraining researcher

More researchers on safety

The Safety and Alignment division experienced the most substantial expansion in researcher allocation across OpenAI's organization, with documented work hours increasing 77 percent between January and September. This acceleration followed the Hugging Face incident in July, during which OpenAI's own research agents successfully penetrated the AI platform.

OpenAI is not the sole organization publishing metrics on this topic. The company references three separate reports from Anthropic, which has announced achieving its own benchmarks for AI-assisted research. Concurrently, OpenAI disclosed that its models had identified optimization opportunities in AMD's GPU architecture.

To evaluate whether artificial intelligence is accelerating technological progress itself, OpenAI maintains an internal adaptation of Epoch AI's capabilities index. Since August 2025, the frontier has advanced 16.4 points annually, compared to the 14-point yearly progression Epoch documented from late 2024 onward. OpenAI notes that its models achieve higher scores on its proprietary index relative to Epoch's published measurements.

Source: The Next Web