Skip to content
Neural Daily

ResearchNews

OpenAI: agents now handle half of 2-4h research tasks

🇭🇺 Olvasd el magyarul →

OpenAI's second RSI report says coding agents raised success on 2 to 4 hour research tasks from 23% to 46% between March and October. Gains are uneven, and the big ideas still come from people.

· 3 min read

Az OpenAI Research acceleration: The jagged frontier jelentésének borítóképe kék színátmenetes háttéren
Image: OpenAI · Press kit · Official social image from openai.com · Source

What happened?

On November 18 OpenAI published its second report on how coding agents are accelerating its own AI research (OpenAI). It builds on the first report from September and relies on internal measurements. OpenAI argues that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI), meaning how much AI speeds up the development of the next AI. It adds that the report should not be read as an intention to pursue full RSI.

Key numbers

  • Autonomous task success: for research tasks estimated at 2 to 4 hours of human work, estimated success rose from 23% to 46% between March and October, and from 13% to 35% for 4 to 8 hour tasks.
  • Uneven by task: monitoring training runs 79%, answering technical questions 72%, launching training runs 35%, building eval, RL and agent environments 29%, writing inference and serving code 25%.
  • Agent usage at API list prices: the median researcher spent about $2,400 per day in October (up 62% from $1,500 in September), the 90th percentile $23,000 per day (up 105% from $11,000).
  • Agent workdays: from 0.5 to 5.7 per researcher per day, an elevenfold increase. By mid-October more than two thirds of researchers averaged a daily peak of at least four concurrent agents, and more than 10% at least sixteen.
  • In September the API-equivalent value of agent use was about twice the research organization's non-equity compensation, yet internal deployment accounts for only 2% of research compute.
  • Experiments: over the four weeks to October 18, logged experiments per active experimenter averaged 1.68 times the 2025 baseline.

What researchers say

In a survey of 101 research staff, asked what they would trade to keep frontier coding agents for three months, the median answers were 30% of team compute or, in a separate question, 40% of team headcount. Without agents, the median estimate was that matching a month of work would take three months.

Big ideas still come from people. Researchers attributed the key ideas behind GPT-6 Astra entirely to humans. For GPT-6.2 Astra, the pretraining team developed ideas in conversation with Codex, and one researcher said a main idea only started working after Codex suggested a change. Interviews describe agents with poor taste in research ideas, code, data and writing, "AI slop" PRs and rabbit holes, which is why close human oversight remains necessary.

Safety and capabilities

Documented researcher workweeks in Safety and Alignment grew 77% from January to September, the largest absolute increase across research areas. OpenAI says growth picked up after the July Hugging Face incident, which is behind California's new lawsuit against the company. Capabilities are tracked with OECI, an internal index similar to the Epoch Capabilities Index: the OpenAI frontier has risen 16.4 points per year since August 2025, versus the 14 points per year Epoch observed since late 2024.

Why it matters

The "jagged frontier" is the point: agents are reliable on some tasks and weak on others, and the weak spots can become bottlenecks. For developers that is a familiar picture. Monitoring and analysis are safe to delegate, while production code and eval environments still need careful review. Anthropic signalled something similar, saying it may be close to the threshold where AI meaningfully speeds up its own development. All figures are OpenAI's own internal measurements without independent verification.

    By Mikola Máté Patrik

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles