
OpenAI research acceleration update: the jagged frontier

OpenAI published an update on how coding agents are accelerating its AI research, extending its September report with new measures of research inputs, the research process, and model outputs. The report is framed as transparency about indicators of recursive self-improvement (RSI) and explicitly says it should not be read as an intention to pursue full RSI; the authors argue AI labs should be required to track and publicly report such indicators to inform governance and safety discussions.
On inputs, agent use inside the research organization has kept rising exponentially. At API list prices, the median researcher now uses about $2,400 per day in agentic AI tools and the 90th-percentile researcher about $23,000 per day, up 62% and 105% from $1,500 and $11,000 in September. Agent workdays per researcher per day rose from 0.5 to 5.7, an elevenfold increase, with pretraining and posttraining teams averaging roughly eight agent workdays per researcher per calendar day and scaling and infrastructure teams roughly four. By mid-October, more than two-thirds of researchers averaged a daily peak of at least four concurrent agents, and over 10% averaged at least sixteen. API-equivalent agentic AI spending in the research organization was about twice non-equity compensation as of September 2026, although equity still accounts for 70–80% of target researcher compensation and internal deployment is only 2% of total research compute.
For the research process, time allocation was estimated from company-public Slack posts, pull requests, and internal documents classified by research area. Safety & Alignment saw the largest absolute increase in documented workweek contributions, almost doubling since the start of 2026, with an inflection after the July Hugging Face incident and subsequent work on safety cases. On delegated tasks, zero-intervention success from August 1 to October 18, 2026 was higher for Communicate (69%) and Analyze (58%) than for Build (43%); task-level examples ranged from monitoring runs for issues (79%) and answering technical questions (72%) to writing code for inference and serving stacks (25%) and building evaluation, RL, and agent environments (29%). Surveys of 101 research staff found a median willingness to trade 40% of team headcount, 30% of current team compute, or 50% of team compute if asked in January 2027 to restore frontier coding agents, and researchers said their teams would need a median of three months to reproduce a month of work with only Copilot and ChatGPT. In a later survey, 27 of 28 respondents chose or leaned toward the latest internal model over a mid-career human engineer with a 2024 model. On idea generation, posttraining and safety researchers attributed all important named ideas to people for both GPT-6 Astra and GPT-6.2 Astra; pretraining researchers attributed GPT-6 Astra ideas to people, but for GPT-6.2 Astra they described developing ideas in conversation with Codex, with one crediting Codex for key insights and another saying the idea worked only after Codex suggested a change. Interviews with ten researchers surfaced five themes: agents help implement, debug, and oversee experiments; agents have poor taste in research ideas, code quality, data quality, and writing; workflows require close human involvement to avoid wasted cycles; researchers expect significant improvements in current bottleneck areas; and scanning or generating content at scale provides large uplift.
For outputs, lines of code changed per active contributor continued rising sharply in 2026 relative to the pre-2025 average, and logged experiments per active experimenter averaged 1.68 times the 2025 baseline for the four weeks ending October 18. OpenAI also introduced OpenAI ECI (OECI), an internal version of Epoch’s Capabilities Index based on public and internal benchmarks. The OECI frontier for OpenAI models increased 16.4 points per year since August 2025, versus Epoch’s observed 14 points per year since late 2024; the report notes that OpenAI models score higher relative to competitors on OECI than on ECI, likely due to differences in the benchmark mixture.
The report closes by emphasizing that the uplift is jagged, not uniform: agents remain weak at several tasks, and close human involvement is still required. It cautions that historical metrics involving human-AI collaboration may not predict future autonomous progress, especially after sudden rapid breakthroughs in mathematics, and says OpenAI will continue refining metrics and expanding reporting on differential acceleration of capabilities versus safety, leading indicators, and capability trends.

