OpenAI estimates coding agents finish 46% of its two-to-four-hour research tasks
Internal measurements show researchers running more agents in parallel, but success depends heavily on the job—and changing delegation habits complicate the comparison.
Loading page…
Internal measurements show researchers running more agents in parallel, but success depends heavily on the job—and changing delegation habits complicate the comparison.
Listen to this story
OpenAI’s November 18 report finds that agents are becoming a valued research input: in October, median daily use was worth about $2,400 at API list prices, 62% above September. But performance varies sharply by task, and OpenAI cautions that higher completion rates may partly reflect researchers choosing better-suited jobs to delegate. The company urges AI labs to track and publicly report indicators of recursive self-improvement, while stressing that this call does not signal an intention to pursue it fully.
Estimated success without human intervention rose from 23% to 46% for two-to-four-hour tasks and from 13% to 35% for four-to-eight-hour tasks between March and October.
From August 1 to October 18, agents succeeded on 79% of model-training monitoring tasks, versus 35% of training launches and 25% of inference and serving-code tasks.
Documented Safety & Alignment workweeks increased 77% from January through September; OpenAI attributes the rise to both hiring and researchers shifting focus, not solely to department growth.
Coding agents inside OpenAI are much better at watching an AI training run than writing the software that serves a model. In a report published November 18, 2026, the company traces rising use and improving task completion—but also sharp limits on which research jobs agents can finish without a person stepping in.
The report, Research acceleration: The jagged frontier, updates OpenAI’s September account of AI-assisted research. It examines three stages: the resources going into research, how researchers organize their work, and the outputs. Its findings combine internal activity measurements with staff surveys and interviews.
One change concerns where people directed their effort. From January through September, documented Safety & Alignment workweeks increased 77%, the largest absolute increase among the research areas measured. OpenAI attributes that growth to both new hires and existing researchers shifting their attention toward safety projects.
That is an estimate of documented attention, not a count of employees assigned to a department. OpenAI classified contributions from company-visible Slack posts, code-change requests and internal documents, then divided each researcher’s week equally across their verified projects. It says safety-work growth picked up following its response to the July Hugging Face incident and subsequent work on safety cases for frontier training.
A survey of 101 research staff, conducted from late August to early September, tested how much researchers valued access. It asked what they would trade to restore coding agents better than GPT-5.2 if access disappeared for three months. In separate questions, the median answers were 30% of team compute or 40% of team headcount. These were hypothetical tradeoffs, not announced staffing decisions.
By October, the median researcher’s daily agent use was worth about $2,400 at published API prices, up 62% from September. At the 90th percentile, usage reached about $23,000, up 105%. Those figures value usage at list prices; they do not represent an internal cash bill. OpenAI says the growth reflects more activity, shifts to more expensive models and changes in how agents behave and are configured.
Parallel work was widespread by mid-October: more than two-thirds of included researchers averaged a daily peak of at least four agents running at once. More than 10% averaged a peak of at least sixteen.
The improvement over time needs one important qualification. Researchers may have changed which jobs they delegated within each estimated-duration group. OpenAI says the rising success rates may therefore reflect both stronger agents and better choices about when and how to use them.
The task-level breakdown makes the unevenness concrete. Between August 1 and October 18, estimated success without human intervention was:
OpenAI warns that weaker task categories could become bottlenecks to overall research gains. It is calling for AI labs to track and publicly report indicators of recursive self-improvement: AI accelerating the development of future AI systems. The company says this disclosure should not be read as an intention to pursue that process fully.
Loading discussion...
Join the conversation
Explain what you would gain—or miss most—in that trade.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.