OpenAI: GPT-5.6 Sol Ran 40 Quantum Measurements at MIT Lab

OpenAI says its GPT-5.6 Sol model, driven through Codex, completed 40 target measurements on a six-qubit superconducting chip in an MIT lab, with a researcher stepping in on only four. The chip lives in the Engineering Quantum Systems Group (EQuS), and the graduate student running the experiment, Beatriz Yankelevich, used the agent to select measurement settings, operate the lab hardware, analyze data and calibrate qubit frequencies, control pulses and information-retention times.

In the OpenAI post, Yankelevich describes the shift in how her days look: "I can have multiple agents working on different problems at once, and I spend most of my time on higher-level work—interpreting results, devising experiments, planning next steps for the agents, reading, and writing."

The framing is narrower than "autonomous physics." The chip is one of a standard type EQuS uses to benchmark its own fabrication, and the workflow the agent ran is one the group already understands. Where the signals got difficult, the agent slowed down: it "performed well with clear signals but sometimes required expert guidance when results were weak, noisy or physically ambiguous," per the writeup. The Quantum Insider covered the same claims the same day.

OpenAI is the only source. There is no peer-reviewed paper attached to the post, and the four interventions are not itemized — we don't know whether the agent misread a hardware fault, picked a bad setting, or simply got stuck. It joins a running tally of agent stories we're tracking this quarter (416 in the last 90 days), most of them in software rather than in a dilution refrigerator.