Research acceleration: The jagged frontier

OpenAI
OpenAI reports that coding agents are increasingly accelerating AI research unevenly across domains, with humans still essential for oversight and idea generation.

Summary

OpenAI provides an update on research acceleration, measuring how coding agents are changing AI development. The report tracks inputs (agent usage, spending, and researcher time allocation), process (tasks delegated to models, success rates, and workflow changes via surveys and interviews), and outputs (lines of code, experiment counts, and the OpenAI Epoch Capabilities Index). Agent usage has grown exponentially, with the median researcher spending about $2,400 per day on agentic tools and agent workdays rising elevenfold. Success rates vary by task category, with communication and analysis tasks succeeding more often than building tasks. Researchers report significant value from coding agents but note limitations in research taste, code quality, and data quality. Safety & Alignment research saw the largest absolute increase in researcher time, partly following the Hugging Face incident. Outputs show sharp increases in code changes and experiment counts, while OpenAI ECI has risen 16.4 points per year. The report emphasizes transparency about recursive self-improvement indicators and the need to maintain human oversight as AI capabilities advance.

(Source:OpenAI)