On November 18 OpenAI published "Research acceleration: The jagged frontier", using internal data to show how coding agents are changing AI R&D. From March to October, agents' estimated solo success rate rose from 23% to 46% on research tasks taking humans two to four hours, and from 13% to 35% on four-to-eight-hour tasks, with wide variation by task (79% for monitoring training runs, 35% for launching them, 25% for inference and serving code). At API list prices, the median researcher now uses $2,400 of agentic AI a day, up 62% from September. Documented Safety and Alignment workweeks grew 77% from January to September. OpenAI says AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI).

Key Takeaways

  • ✓Task success: from 23% to 46% on two-to-four-hour tasks and from 13% to 35% on four-to-eight-hour tasks (March to October); 79% for monitoring training runs, 35% for launching them and 25% for inference and serving code
  • ✓Usage: the median researcher spends about $2,400 a day on agents (up 62% from September) and the 90th percentile about $23,000 (up 105%); agent workdays per researcher rose from 0.5 to 5.7
  • ✓Concurrency: 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
  • ✓Cost mix: in September the API-equivalent value of agent use was about twice non-equity researcher compensation, yet internal deployment used only 2% of research compute
  • ✓Safety: documented Safety and Alignment workweeks grew 77% from January to September, the largest absolute increase of any research area; OpenAI argues labs should be required to report RSI indicators
OpenAI Publishes Its Second Research Acceleration Report: Agents Now Complete 46% of Two-to-Four-Hour Research Tasks Alone, Median Researcher Uses $2,400 of Agents a Day, Safety and Alignment Effort Up 77%
🖼️Official Media
Click to view high-res
⚡

Key Decision Metrics at a Glance

TL;DR Verdict
🏆 Benchmark RankRank
No model benchmark in this story · see current rankings
Full Leaderboard→
🛠️ Ecosystem & IDERules
OpenAI · research acceleration · recursive self-improvement
Setup IDE Rules↗
🧭

Turn your technical choice into a development budget

Compare 40 dev plans & simulate token costs vs $20/mo subscriptions

🔬

In-Depth Technical Analysis

Background

Whether AI can speed up its own development, known as recursive self-improvement (RSI), is central to governance and safety debates, yet outsiders rarely see internal lab data. OpenAI published its first research acceleration report in September and released a second one, "Research acceleration: The jagged frontier", on November 18 with more quantitative and qualitative measures.

How it works

The report is organized around inputs, the research process and outputs: inputs track coding-agent usage, the process looks at how much research work agents complete alone and how humans stay involved, and outputs look at capability trends. OpenAI measures usage in "agent workdays", where every eight hours of agent runtime counts as one, and in spend at API list prices.

Numbers that matter

Per OpenAI's report, from March to October success on two-to-four-hour tasks rose from 23% to 46% and on four-to-eight-hour tasks from 13% to 35%. The gap by task is large, with 79% for monitoring training runs but only 25% for inference and serving code, hence the "jagged frontier". The median researcher spends about $2,400 a day on agents, agent workdays per researcher rose from 0.5 to 5.7, and Safety and Alignment workweeks grew 77%.

What it means for developers

For engineering teams, the report offers a useful baseline: agents already handle long-running work such as monitoring well, while complex systems code still needs humans in charge. It also shows that heavy users increasingly run many agents at once, so concurrency and subagents are worth trying in your own workflows.

Action HubReady to adopt this in production?

Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.

Up NextX Debate: AI Commentator Chubby Says Claude Code and Opus 5.5 Have Been His Daily Driver for Two Months and GPT Is "Not Close," but Concedes Computer Use Is Still GPT's Edge; 150+ Replies Focus on Splitting Work by Task and the Same-Day Claude Login Outage
Read Now→