OpenAI is showcasing an unusual research application for its agentic models: running physics experiments. The experiments ran on superconducting qubit chips, with MIT’s Beatriz Yankelevich, a graduate student in the Engineering Quantum Systems group, linking GPT-5.6 Sol through Codex to the coordinating software.
After fabrication, packaging and cooling in a dilution refrigerator, a qubit chip is reachable only through software, so the whole workflow suits an AI agent. She handed Codex measurement-level skills describing how each experiment runs and gets judged, then turned it loose on an uncalibrated six-qubit chip that benchmarks the group’s fabrication work.
The agent chose measurement parameters, operated the hardware, analyzed the returning data and decided whether to refine a measurement or move on. Faced with clean signals, the agent ran standard calibration sequences nearly hands-off: locating each qubit’s transition frequencies, adjusting the control and readout pulses, and timing how long quantum information survived. Yankelevich said the setup saved significant time and let experiments run without constant supervision, freeing her for experiment design and data analysis.
The limits were equally instructive. Weak or noisy signals slowed GPT-5.6 Sol’s path to good parameters, and occasionally it needed a human prompt. For OpenAI, the demo shows agents can take charge of clearly-scoped experimental workflows today, while interpreting messy physical results remains researcher territory.
