Codex and GPT-5.6 Sol automate qubit calibration in an MIT quantum computing lab

OpenAI has published a new case study, dated September 8, 2026, on the use of its GPT‑5.6 Sol model — connected to Codex, its coding agent — to run and refine quantum computing experiments in a university lab. The protagonist is Beatriz Yankelevich, a graduate student in MIT's Engineering Quantum Systems Group (EQuS), who investigated whether an AI agent could take on part of the routine work of qubit calibration in order to free up time for experiment design and data analysis.

Yankelevich's group works with superconducting qubits, the basic units of quantum information, which are cooled to near absolute zero inside dilution refrigerators and controlled with microwave pulses. Unlike conventional processors, these chips can only be manipulated through software once they have been fabricated, packaged and cooled, which makes them natural terrain for an AI agent to operate autonomously: all the experimental work consists of sending instructions, reading signals and deciding the next step.

As the article describes, calibrating a quantum chip requires a long chain of interdependent measurements — from identifying each qubit's transition frequencies to calibrating the control and readout pulses, or determining how long quantum information is preserved — where each result conditions the next measurement. In addition, qubit properties can drift over time and produce unexpected behavior, something that normally requires an experienced researcher's intuition to interpret correctly.

Yankelevich put GPT‑5.6 Sol to the test on an uncalibrated six-qubit chip, of a standard type the EQuS group routinely uses to assess its fabrication process. Through Codex, she provided the agent with a series of 'skills', or specific instructions on how to run and evaluate each type of measurement, along with the chip's design parameters. From there, the model chose the measurement parameters, operated the lab hardware, analyzed the resulting data and decided whether to repeat the measurement with adjustments or accept it as valid and move on to the next stage.

The case study notes that when signals were clear, Codex completed the entire standard calibration sequence almost autonomously: frequency identification, tuning of control and readout pulses, and measurement of quantum coherence times. However, the article itself acknowledges a relevant limitation: when experimental signals were weak or noisy, the model took longer to find suitable parameters and at times needed the intervention of an experienced researcher. OpenAI sums up this observation by saying that today's agents can handle clearly defined experimental workflows well, but that interpreting ambiguous physical results remains a challenge.

Beyond the specific experiment, the article indicates that the EQuS group has made routine use of agents for standard chip characterization measurements — a process that, as explained, can take a researcher several days per chip. Yankelevich is quoted explaining that she can leave agents running measurements for hours, overnight or while she works in the clean room, and supervise or redirect the work remotely from her phone. For less routine experiments, she says she assigns agents narrower goals and takes advantage of their ability to write, modify and test control, analysis and simulation code, so that reviews are checked directly against real lab measurements.

The researcher adds that she has built her own infrastructure to guide agents across different parts of her work — measurement, theory and chip design — and that she can now have several agents working in parallel on different problems while she concentrates her time on higher-level tasks: interpreting results, designing experiments, planning next steps, reading and writing.

It is worth noting that this is a case study published and promoted by OpenAI itself, in line with other similar pieces the company has put out about the use of its models in scientific research (biology, immunology). As in those cases, the text offers no quantitative data — success rates, hours saved or formal comparisons against manual work, for example — but rather testimony and descriptive examples from the research team itself, so it is best read as a piece of corporate communication illustrating a specific use case rather than as an independent scientific study of AI agents' capabilities in lab settings.