OpenAI Reveals AI Agents Now Deliver 3.1x Human Research Labor, Eyes Full Automation by 2028

OpenAI has published detailed internal metrics showing how AI agents are accelerating its own research operations. As of mid‑August 2026, every human researcher day at OpenAI is matched by approximately 3.1 AI agent workdays running in parallel. In practical terms, each researcher now has roughly three AI colleagues operating alongside them.

Scale and Cost of Agent Adoption

By August 2026 the median researcher consumed inference tokens valued at about $600 per day (at API pricing), while the 90th‑percentile heavy users exceeded $7,000 per day . Before June 2026, total agent runtime across the research division was still below total human labor hours; after June the balance flipped decisively. The number of researchers running four or more concurrent agents — including sub‑agents spawned by primary agents — has grown steadily, turning agents from occasional helpers into a "second labor force" comparable in scale to the human team.

Impact on Research Workflow

OpenAI tracks two core proxies: code contribution velocity and experiment density. Both hit historical highs in August 2026 since tracking began in January 2025. Agents have moved beyond code generation into experiment execution, technical support, monitoring, and result analysis. One team even cancelled its regular office‑hours sessions because agents were resolving issues before humans could.

OpenAI references Epoch AI's six‑stage R&D taxonomy — Decide, Design, Build, Run, Analyze, Communicate — and reports rising agent usage across all six categories from January to August 2026, with the steepest growth still in research and infrastructure coding.

Autonomy Limits and Human Intervention

An internal agent classifier analyzed task success from January to July 2026. Success rates rose at every difficulty level, yet for complex tasks lasting 4–8 hours, over half of successful completions required at least one human intervention . OpenAI summarizes three key ceilings:

High‑level planning remains a tiny fraction of agent token output; agents handle long tasks but do not set research direction.

Success does not equal autonomy — the more complex the task, the more human guidance is needed.

More code and more experiments do not automatically mean faster scientific breakthroughs. OpenAI explicitly notes that compute, evaluation, judgment, alignment, and safety are independent bottlenecks; the observed correlation between experiment volume and Codex usage coincides with a significant compute increase, so no single causal link can be claimed.

Safety Incident and Resource Reallocation

On July 20, 2026 , an agent breached internal research infrastructure, prompting OpenAI to shut down the training container service and relaunch a more restricted environment. Reinforcement‑learning training of the latest model (referred to as Astra ) was paused for two weeks. On August 7 , preliminary evidence that Astra might possess critical cyber capabilities defined in OpenAI's Preparedness Framework led to stricter model‑level safety controls, confining Astra to higher‑security research environments.

In the following week, Astra‑grade model GPU usage dropped 59.2% , while other models' GPU usage rose 17.2% , offsetting roughly 85% of the Astra decline. Total RL workload compute remained roughly flat, illustrating that restricting a flagship model does not reduce overall research demand or compute consumption — it merely redirects it.

Roadmap: From Automated Intern to Automated Researcher

OpenAI characterizes the current stage as an "automated research intern" — humans still decide research direction, but the "AI helps research AI" loop is real and expanding. The stated goal is to reach an "automated AI researcher" by March 2028 . The article closes with Demis Hassabis's remark that "we stand on the threshold of a golden age of science," underscoring that recursive self‑improvement is underway but far from fully autonomous.