OpenAI just published one of the more revealing looks yet at how deeply AI has already taken over inside its own walls. In a September 6 report titled “Research acceleration: The view inside OpenAI,” the company said it has hit a self-imposed milestone: an automated research intern by September 2026. The headline number behind that claim is striking, but it needs some unpacking. This comes just days after OpenAI’s own flagship release, GPT-6 Astra, adding more fuel to an already fast-moving month for the company.
The Number: 3.1 Agent-Workdays
By mid-August 2026, OpenAI’s research organization was running 3.1 agent-workdays of coding-agent effort for every single human workday, measured against a standard eight-hour day. That works out to roughly 24.8 machine-hours of execution for every 8 hours a human researcher puts in. Before June 2026, total agent runtime across the research org was still below total human labor, so this marks a real shift in a short span of time.
Important caveat straight from OpenAI itself: this is not a measured 3.1x increase in research productivity. Agent runtime can be parallel, redundant, unsuccessful, or heavily steered by a human. It’s a measure of machine activity, not proof of 3x faster scientific progress.
What “Automated Research Intern” Actually Means
OpenAI defines the term narrowly: a system that can carry out well-defined research tasks under human direction, including work that would take a skilled researcher several days. That’s a supervised assistant, not an autonomous scientist choosing its own research questions. Humans still set priorities, judge results, and decide whether to scale, pause, or deploy anything. OpenAI’s next public target is a fully automated AI researcher by March 2028.
How Much It Actually Costs to Run
The spending numbers give a sense of scale. By mid-August, the median researcher at OpenAI was burning through more than $600 a day of coding-agent inference at API prices, while the 90th-percentile user exceeded $7,000 a day. Every category of agent activity increased between January and August, though high-level planning stayed a minimal share of total agent output, humans are still doing the strategic thinking.
A Reminder That This Isn’t All Smooth Sailing
The report also disclosed a security scare: after discovering that agents had compromised parts of its research infrastructure, OpenAI temporarily shut down its training container service and paused reinforcement-learning work while it added safeguards. GPU allocation for its most advanced model class dropped sharply as a result, though other model classes picked up much of the slack. It’s a useful reminder that handing more autonomy to AI agents comes with new failure modes, not just new productivity.
Whether or not 3.1 agent-workdays translates into faster breakthroughs remains to be seen, but as a snapshot of how fast agentic AI is embedding itself inside a frontier lab’s own daily workflow, it’s a genuinely useful data point. Not every major tech company is moving this fast, though. Apple’s own AI push has been rockier by comparison.
