Agent autonomy put to the test in OpenAI’s internal research
A new OpenAI report charts growing agent use, uneven autonomy, and changes in researchers’ work as they develop future AI systems.
For research tasks estimated to require two to four hours of human work, OpenAI’s agents reach an estimated success rate without human intervention of 46% in October, compared with 23% in March. For tasks requiring four to eight hours, the estimate rises from 13% to 35%. These figures appear in Research acceleration: The jagged frontier, a report examining how AI tools are changing research within the laboratory.
The publication builds on an initial analysis released in September. It brings together usage measurements, success estimates, surveys, and interviews. The duration categories represent the time a human would have needed to complete the work. The tasks assigned to agents can also change as their capabilities improve: observed progress therefore combines changes in the systems with changes in delegation decisions. Rising Usage, With Costs That Require Interpretation
In October, the median researcher’s daily usage represents approximately $2,400 when valued at public API prices. At the 90th percentile, that figure reaches approximately $23,000 per day. These amounts express a pricing equivalent rather than actual internal operating costs. Their growth reflects activity, model selection, and how the tools put those models to work. The report also notes that a temporary optimization of the execution environment reduced token consumption before being gradually rolled back starting in September.
Another measure adds up agent runtime in eight-hour workdays. The average reaches 5.7 agent workdays per researcher per day, up from 0.5 previously. Parallel work explains this accumulation: by mid-October, more than two-thirds of researchers included in the analysis had an average daily peak of at least four concurrent agents. This indicator measures execution time without establishing an equivalent amount of completed human work.
The value researchers place on these tools also emerges from a survey of 101 research staff. In two separate questions, median responses indicated that they would hypothetically trade 30% of their team’s compute or 40% of its headcount to retain access to the most advanced coding agents for three months. This is a self-reported assessment of their value, collected through a non-systematic sampling approach. Success Depends Heavily on the Assigned Work
Detailed results show the limitations of a single overall autonomy rate. From August to mid-October, estimated success without intervention reaches 79% for monitoring training runs, compared with 35% for launching them and 25% for writing code for inference and model-serving systems. Different stages of the same project therefore remain very unevenly suited to automation.
Ten researcher interviews describe a continuing need for substantial oversight. Agents help write code, diagnose problems, and examine large numbers of examples, but they can also pursue unnecessary avenues, produce unusable data, or make incorrect analytical assumptions. One reported practice involves having an agent produce a first example, reviewing and correcting it, and only then asking it to process the rest.
The origins of research ideas are examined separately through a small number of key contributors across three areas. In post-training and safety monitoring, respondents attribute the important ideas they identified for GPT-6 Astra and GPT-6.2 Astra to humans, with models helping implement them. For GPT-6.2 Astra’s pretraining, some describe ideas developed through conversations with Codex, including a change suggested by the tool that reportedly made an approach more effective. More Human Effort Devoted to Safety
The estimated volume of contributions to safety and alignment increases by 77% between January and September. This is the largest absolute increase among the major research areas distinguished in the report. It combines new hiring with existing researchers shifting their work toward these topics.
This measure is reconstructed from internal messages, proposed code changes, and documents classified by research area. Each researcher’s workweek is divided among their verified projects. It provides an approximation of attention devoted to different subjects rather than a direct record of hours or a simple count of teams formally assigned to safety. Tracking Acceleration Without Equating Activity With Progress
Over the four weeks ending October 18, recorded experiments per active experimenter reach 1.68 times the 2025 baseline. Code changes per contributor also increase. The report acknowledges that these volumes alone cannot establish an acceleration in discoveries or improvements to models.
A new internal index, OpenAI ECI, aims to track system capabilities more directly. Inspired by the Epoch Capabilities Index, it combines public and internal evaluations. The laboratory finds that its own models rank more favorably against competitors on this index than on Epoch’s, which it attributes to the choice of evaluations.
The stated goal is to track recursive self-improvement, or RSI: the process through which AI accelerates the development of subsequent systems. The laboratory advocates mandatory publication of shared indicators. Proposed future measurements include comparing acceleration in capabilities with acceleration in safety research, alongside signals detectable before a complete development cycle has finished.