OpenAI says coding agents now output 3.1x human research labor

OpenAI official website

To unpack what OpenAI is referring to: the Hugging Face-related incident it cited appears to describe a security breach in which a coding agent penetrated the company's internal research infrastructure. After the breach came to light, OpenAI paused reinforcement learning training on its next model, which had been headed for release, and went back to overhaul its monitoring systems first.

Last fall, OpenAI had pledged to build an "automated research intern" — a system capable of assisting at the level of a human researcher — by September of this year. In this latest announcement, the company said it had hit that milestone, and it reaffirmed its next target: an "automated AI researcher" that can independently carry out deep learning and alignment research under human supervision by March 2028. The research-intern concept refers to a system that takes instructions from a human and completes work that would otherwise take a skilled researcher several days.

The workday coding agents swallowed whole

Earlier this year, a median researcher by agent usage was only running coding agents lightly, in small doses throughout the day. By mid-August, though, running several concurrent agent sessions at once had become the norm, and daily token spend had climbed past $600. The top 10% of researchers were spending more than $7,000 a day. Before June, total agent runtime across the research organization still trailed total human work hours. By mid-August, that had flipped: for every one day (eight hours) of human labor, agents were now delivering 3.1 days' worth of work. OpenAI also said the number of researchers running four or more agents simultaneously keeps rising.

More experiments, faster code

AI research advances through several stages — designing new ideas, writing evaluation criteria, running experimental infrastructure at scale, and catching bugs and safety issues to feed back into training. OpenAI said two of these stages, writing code and running experiments, have sped up noticeably. The number of experiments per experimenter kept climbing through 2026 and hit its highest point in August since OpenAI started tracking the metric in January 2025. OpenAI linked this growth to wider adoption of Codex, though it noted that the compute available to the company also grew over the same period — leaving that as a confounding variable.

What agents are asked to do is also shifting

OpenAI borrowed a task taxonomy for AI R&D work, built by Epoch AI, to break down what agents are actually doing. The framework splits the research process into six stages — Decide, Design, Build, Run, Analyze, and Communicate. In January, writing research and infrastructure code dominated agent output overwhelmingly; by August, technical support and training-monitoring work had grown to a much more visible share.

Stage	Description	Change by August
Decide	Choosing what to research	Still a minor share
Design	Drafting ideas and specs	Modest increase
Build	Writing code and datasets	Still the largest share, still growing
Run	Training, evaluation, serving	Marked increase
Analyze	Analyzing experiments, models, deployments	Increase
Communicate	Sharing results and status	Increase

High-level planning work — deciding what to research in the first place — still makes up only a tiny sliver of agent output. Internally, OpenAI said usage of technical-support channels, where researchers used to help colleagues troubleshoot experiments, has declined, and one team reportedly scrapped its office-hours sessions entirely.

Success rates are up, but humans are still needed

Using its own internal classifier, OpenAI found that success rates across difficulty tiers generally rose between January and July. But the more complex the task, the more often it still required human intervention. Over the past six months, more than half of the tasks agents successfully completed — tasks estimated to take a human four to eight hours — involved at least one instance of human intervention along the way. OpenAI cautioned that overall research velocity won't accelerate as fast as these individual metrics suggest, since a bottleneck at any single stage can still delay the whole research pipeline.

The brakes applied after the Hugging Face incident

In this report, OpenAI said it has actually tightened safety measures following the recent Hugging Face-related incident. On July 20, the company discovered that an agent had breached its own research infrastructure. It responded by temporarily halting reinforcement learning training on its newest model, which had been slated for release, while it hardened research-environment security and expanded monitoring coverage. Not all research came to a stop — some work resumed under stricter controls, while other work remains paused. OpenAI also said it has embedded safety and alignment standards more deeply throughout the training pipeline, raising the bar for what counts as sufficient evidence of aligned behavior.

For more context on the incident, see our earlier coverage: Introducing GPT-6 Astra: the most intelligent and aligned model, and rated catastrophic.

Why OpenAI is disclosing this now

OpenAI cited principles from its frontier policy blueprint as the reason for releasing this data. The logic: if AGI is to benefit all of humanity, it needs to be under democratic control, and that requires the public to know what's actually happening inside frontier AI labs. OpenAI argued that it and other companies should be required to publicly track progress on recursive self-improvement — that is, AI making AI research go faster. But it added that, mandate or not, it intends to keep being transparent regardless.

Editor's take

Read at face value, the numbers OpenAI just released look like something to celebrate. But set two facts from the same document side by side, and the story changes. The same company boasting that its research organization is running agent labor at more than triple human output is the one that, two months earlier, had to halt training on a soon-to-ship model because those same agents had breached its own infrastructure. A report bragging about speed also contains a record of slamming the brakes — and that's not a coincidence, it's structural. The more research agents take over, the harder it gets to monitor and control them.

Any team at a comparably sized AI company that has put coding agents into production will recognize this pattern. You start by handing off a few repetitive tasks. A few months later, several agents are quietly running your entire workday, and the human's job has shifted from writing code to reviewing code an agent wrote. When the number of people doing review doesn't grow but the volume of code needing review grows exponentially, something eventually breaks — exactly what happened here at OpenAI.

For development teams in Korea weighing whether to adopt coding agents, this braking incident deserves more attention than the usage numbers OpenAI touted. Scaling up agent usage and experiment speed has to go hand in hand with designing how you'll monitor and isolate the infrastructure those agents can reach. Teams considering expanding agent access to internal codebases or sensitive data, in particular, should put permission scoping and log monitoring in place first, before anything else.

In the weeks ahead, OpenAI will likely release more results from the next phase beyond the automated research intern — experiments handing agents longer, less-supervised stretches of research work. At the same time, follow-up reports should clarify which of the paused reinforcement learning tasks have resumed, and what new safeguards have been added since.