Six months from now, OpenAI says it will have an AI system capable of autonomously running research experiments, analyzing data, and making small scientific discoveries. That was the promise CEO Sam Altman made during an October 2025 livestream, and the clock is ticking.
MIT Technology Review reports that OpenAI has refocused its entire research operation around this goal, redirecting resources toward what the company calls its "AI researcher" project. The plan has two phases: an automated research intern by September 2026, running on hundreds of thousands of GPUs, followed by a fully autonomous AI researcher by March 2028.
Where Things Actually Stand
The building blocks are visible in OpenAI's recent product launches. Deep Research, which shipped in February 2025, handles web browsing and data synthesis. ChatGPT Agent, launched last July, combines web interaction with conversational intelligence. GPT-5.4, which dropped on March 4, scored 75% on OSWorld-Verified (a benchmark for computer use tasks), edging past the human baseline of 72.4%.
But there is a telling gap. On OpenAI's own FrontierScience benchmark, which tests scientific reasoning across physics, chemistry, and biology, GPT-5.2 scored 77% on structured problem-solving but only 25% on open-ended research questions. That 25% number is the honest signal here. Solving well-defined problems is a different skill than forming hypotheses and designing experiments from scratch.
Chief Scientist Jakub Pachocki has acknowledged that current models handle tasks with roughly a five-hour time horizon. For real research breakthroughs, he has said it would be "worth dedicating entire data centers' worth of computing power to a single problem."
The Resource Question
OpenAI initially committed $1.4 trillion over eight years for compute infrastructure. That number has since been revised down to roughly $600 billion through 2030, a 57% cut. The company has also restructured from a nonprofit to a public benefit corporation, specifically to attract the capital this kind of project demands.
The corporate restructuring, the AMD partnerships, the Stargate Project data centers, the $38 billion Amazon deal - all of it feeds into this single bet. OpenAI is not hedging.
The Gap Between Intern and Researcher
Altman has been unusually candid about the uncertainty. "We may totally fail at this goal," he said, "but given the extraordinary potential impacts we think it is in the public interest to be transparent about this."
The September 2026 intern is expected to handle end-to-end work packages: literature reviews, data analysis, running simulations, proposing follow-on experiments. The March 2028 researcher would go further, proposing original hypotheses and designing experimental protocols independently.
Andrej Karpathy, OpenAI co-founder and former Tesla AI director, recently published an open-source tool called autoresearch that lets AI agents run ML experiments overnight on a single GPU. It is 630 lines of code. The gap between that and OpenAI's vision of datacenter-scale autonomous science is enormous, but Karpathy's project shows the basic concept is no longer theoretical.
The real test is not whether OpenAI ships something it can call a research intern by September. It is whether that system produces results that actual scientists find useful. A 25% score on open-ended research reasoning suggests there is still a long way to go, even with hundreds of thousands of GPUs behind it.