OpenAI Reports Uneven Gains in Agent Research Success
OpenAI reports estimated agent success on sampled two-to-four-hour research assignments rising from 23% in March to 46% in early October. Results vary by task, from 79% estimated success in monitoring to 25% in inference/serving code; the company’s internal measurements do not establish broad autonomy and need independent comparison.
Sources (2)
Updated Nov 21, 2026