GPT-5.6 Sol runs autonomous qubit-calibration experiments for an MIT quantum lab

OpenAI says a graduate researcher at MIT's Engineering Quantum Systems Group (EQuS) connected GPT-5.6 Sol, routed through Codex, directly to the lab software controlling a superconducting quantum chip — and had the model run, analyze, and steer real qubit-calibration experiments with minimal supervision.

What's new

• Beatriz Yankelevich, an EQuS graduate student, gave Codex "measurement-specific skills" describing how to run and evaluate each experiment on an uncalibrated six-qubit chip, one of a standard type EQuS uses to benchmark its fabrication process.
• Using those skills and the chip's design targets, GPT-5.6 Sol chose measurement parameters, operated the hardware, analyzed the resulting data, and decided whether to refine the measurement or save the result for the next step.
• On clear signals, it autonomously identified the qubit's transition frequencies, calibrated the microwave pulses used to control and read it, and determined how long it retained quantum information — a standard characterization sequence that normally requires a researcher's continuous attention.
• The model had more trouble when signals were weak or noisy, taking longer to find suitable parameters and sometimes needing an experienced researcher to step in. OpenAI's writeup frames this as the current boundary: agents can handle clearly defined experimental workflows, but ambiguous physical results remain a challenge.
• EQuS says it "now regularly uses agents to handle routine measurements," freeing researchers for higher-level analysis and experiment design.

Context

Preparing and characterizing superconducting qubits — cooled to near absolute zero inside dilution refrigerators and controlled entirely through software once packaged — traditionally takes months and hundreds to thousands of preliminary measurements. Because researchers interact with a fabricated chip purely through code, EQuS's workflow is structurally similar to a software codebase an agent can operate on, which is what made it a natural testbed for Codex rather than requiring any new physical tooling.

Why it matters

This is a concrete public case of a frontier OpenAI model directly operating physical-science lab infrastructure — choosing measurement parameters and steering hardware — rather than just reasoning about the domain in a chat window, extending the "agents doing real research work" narrative OpenAI has also pushed this month around mathematics and its internal coding-research pipeline. As Yankelevich put it: “I can check in from my phone, see what they’ve done, and steer them if something needs fixing or if I want to explore a different direction.”