← Back to feed

OpenAI published a case study Wednesday documenting GPT-5.6 Sol, harnessed to Codex, autonomously calibrating a six-qubit superconducting quantum chip inside MIT’s Engineering Quantum Systems Group (EQuS) — one of the first published accounts of a frontier AI agent operating quantum lab hardware end-to-end.

The Experiment

MIT doctoral researcher Beatriz Yankelevich connected Codex to the lab software controlling a six-qubit chip of the type EQuS routinely uses to benchmark its fabrication process. The chip was uncalibrated: no baseline measurements had been run, no qubit resonance frequencies were known.

Superconducting qubits are cooled to near absolute zero inside dilution refrigerators and controlled via precisely timed microwave pulses. Calibrating them requires a series of interdependent measurements — each result shapes what happens next. Qubit properties can drift unpredictably, and handling drift normally requires an experienced researcher to recognize anomalies in real time and adapt the measurement sequence.

Yankelevich gave Codex measurement-specific skills — structured descriptions of how to run and evaluate each experiment type — alongside the chip’s design targets. From there, GPT-5.6 Sol:

1. Selected initial measurement parameters
2. Operated the lab hardware to execute the measurements
3. Analyzed the resulting microwave signal data
4. Decided whether to refine the current measurement or advance to the next step
5. Iterated through the full calibration sequence for all six qubits

According to the published case study, Sol “often complete[d] routine measurement workflows autonomously,” allowing experiments to run overnight without supervision and freeing Yankelevich to focus on experimental design and result interpretation rather than execution.

Why This Is Harder Than It Looks

Qubit calibration is a genuine test of adaptive decision-making under noisy feedback. The measurements are interdependent: a frequency sweep determines which control parameters to use for subsequent coherence measurements, which in turn constrain gate optimization. Unexpected physical behavior — a qubit resonance shifted by a manufacturing defect, a cable artifact in the signal chain — can cascade into hours of manual troubleshooting.

The fact that Sol handled this on an uncalibrated chip (rather than recalibrating a previously characterized device) means the agent had no prior state to rely on. It was starting from scratch with only the chip’s design targets as a prior.

What It Doesn’t Yet Do

The case study is a controlled demonstration, not a production deployment. Codex was given structured skill definitions for each measurement type — it was not inventing new experimental methods. Yankelevich remained responsible for the higher-level experimental program: which qubits to characterize, what research questions to pursue, and how to interpret results in the context of ongoing EQuS projects.

Still, the scope is notable. Quantum computing research has historically been bottlenecked by the time cost of routine characterization — months of preliminary measurements before a chip is usable for actual research. AI agents that absorb that overhead shift researcher time toward the parts that actually require human judgment.

Broader Context

OpenAI’s Codex crossed 8 million users in July 2026, primarily in software engineering workflows. The MIT case study is part of a deliberate push to document use cases beyond code — scientific instrumentation being the clearest example where agentic AI can operate hardware rather than just generating output.

Quantum computing labs are a structurally good fit: the control interface is entirely software, measurement protocols are well-defined, and the bottleneck is iteration speed rather than physical dexterity. OpenAI has not disclosed which other lab verticals are in testing.