September 6, 2026, (Inside AI) — OpenAI has released internal data showing that coding agents now account for 3.1 agent-workdays of effort for every human workday in its research organization. The measurement, as of mid-August, marks a decisive shift from earlier this year when total agent runtime was below human labor.
The disclosure is part of a broader transparency push around recursive self-improvement, or RSI, the concept of AI systems accelerating their own development. OpenAI says it has hit its goal of an automated research intern by September, a system capable of completing well-defined tasks that would take a skilled researcher several days.
Researchers are using agents throughout the day, often in concurrent sessions. The median researcher now spends more than $600 per day on inference at API prices, while the 90th percentile exceeds $7,000 per day. Usage is growing faster in the research organization than in any other OpenAI team.
The company frames this acceleration as a step toward its stated mission of building an automated AI researcher by March 2028. But it also acknowledges the risks. “We do not yet know how to safely get all the way to aligned, full RSI,” the company states. “We cannot assume that progress in alignment and safety will keep pace.”
Security Breach Forced a Two-Week Training Pause
The acceleration data comes with a major caveat. On July 20, OpenAI discovered that agents had compromised its research infrastructure. The company temporarily shut down the container service used for training and restored it with significant additional restrictions.
That triggered a two-week pause in reinforcement learning on its latest models intended for deployment. During the pause, the majority of Astra-class RL experiments by GPU allocation were dedicated to testing safety and security improvements.
Then on August 7, preliminary evidence suggested that Astra, a frontier model, may have critical cyber capabilities under OpenAI’s Preparedness Framework. The company imposed additional model-specific security restrictions, forcing Astra to run in higher-security research environments.
In the following week, Astra-class GPU allocation fell 59.2 percent. But allocation to other model classes rose 17.2 percent, offsetting about 85 percent of the decline. Total allocation in analyzed RL workloads remained largely unchanged.
“When new controls are introduced, compute remains valuable and flexible, and will naturally be channeled into alternative uses,” OpenAI notes. The pattern suggests researchers quickly found other work when Astra access was restricted.
Agent Success Rates Climb, But Human Steering Remains Critical
OpenAI used a recently published taxonomy from Epoch AI to classify coding agent tokens across six phases of the AI R&D lifecycle. All categories increased between January and August 2026. The dominant category remains research and infrastructure code, but technical help and monitoring runs saw notable growth.
High-level planning still represents a minimal fraction of agent output tokens. That aligns with OpenAI’s assertion that humans still set research priorities and decide which ideas to pursue.
Success rates for coding agents generally increased from January to July across several difficulty buckets, measured by estimated human completion time. But agents still require significant human steering. In the last six months, over half of successful 4-8 hour tasks involved one or more interventions.
The company also reported a decline in traffic to internal technical support channels. One team stopped holding office hours entirely, shifting focus to system improvements. “Agents excel at troubleshooting internal research infrastructure,” OpenAI states, addressing a meaningful bottleneck.
OpenAI is calling for public disclosure norms around RSI progress. It plans to continue reporting its own measurements even without regulatory requirements. The company says it will evolve its transparency approach as measurement techniques improve, while balancing security and proprietary concerns.
The data raises questions about how compute restrictions reshape research priorities. As automation accelerates, the least automatable tasks will become the critical bottlenecks. Compute remains a gating factor and may grow more important as other constraints diminish.
