Codex Joins the Calibration of a Real Quantum Chip: What One Public Run Shows

On 8 September 2026 OpenAI published a case study: Beatriz Yankelevich, a graduate student in MIT's Engineering Quantum Systems Group (EQuS), used GPT-5.6 Sol, harnessed to Codex, in her own experimental workflow. EQuS studies superconducting qubits, which are cooled to near absolute zero in dilution refrigerators and controlled with microwave signals.^[S1]

The premise the article sets out is that once a chip has been fabricated, packaged and cooled, researchers interact with it entirely through software. Connected to the lab software that coordinates experiments, Codex could run measurements, analyse the results and decide what to try next.^[S1]

What the model actually touched

The test object was an uncalibrated six-qubit chip, one of a standard type EQuS routinely uses to benchmark its fabrication process. Yankelevich gave Codex measurement-specific skills explaining how to run and evaluate each experiment. Using those skills and the chip's design targets, GPT-5.6 Sol chose measurement parameters, operated the hardware, analysed the data, and then either refined the measurement or saved the result for the next one.^[S1]

First, what the two measurements are for. A Rabi measurement steps up the drive pulse's amplitude and answers "how strong a pulse flips the qubit from 0 to 1", which sets the π and π/2 pulse amplitudes; a Ramsey measurement puts a variable wait between two π/2 pulses and answers "exactly where the qubit's frequency sits". The run screenshot attached to the article shows the grain of that interaction. Through a Jupyter integration the model issued a 1D Rabi sweep over 0-0.4 V and fit an Xπ amplitude of 50.6 mV and an Xπ/2 amplitude of 24.8 mV (R² 0.994). The Rabi chevron's driven centre, 199.27 MHz, sat about 0.67 MHz above the half-flux spectroscopy minimum, so it widened the Ramsey detuning span to ±2 MHz, ran it on the 4 ns hardware grid with active reset, and stored a bare transition frequency f01 = 198.592407 MHz. The two steps are marked 17 s and 67 s of acquisition plus analysis and about 1k tokens each, the model's own estimate. The model was setting sweep ranges and pulse amplitudes on a real chip and reading back real measured signals, not simulator output.^[S1]

The boundary sits here too. The measurement types themselves (Rabi, Ramsey, spectroscopy, coherence) are standard procedures, and how to run and evaluate them was written into the skills by the researcher. The model chose parameters within those procedures, judged the results and chained the next step; the article does not say it designed new pulse sequences or new experiments.^[S1]

The outcome

When the signals were clear, Codex completed a standard sequence of measurements with little researcher intervention: it identified the qubit's transition frequencies, calibrated the pulses used to control and read it, and determined how long it retained quantum information. When signals were weak or noisy, it took longer to find suitable measurement parameters and sometimes needed guidance from an experienced researcher. The article's conclusion is that current agents can handle clearly defined experimental workflows, while interpreting ambiguous physical results remains a challenge.^[S1]

The article describes qubit calibration as a series of interdependent measurements, each result shaping what happens next; qubit properties can drift and unexpected physical behaviour can produce inconsistent results, which experienced researchers recognise and adapt to. That combination of software control, repeated measurement and adaptive decision-making is what makes it a use case for agents. Preparing and running qubit experiments can take months and hundreds to thousands of preliminary measurements.^[S1]

EQuS fabricates many of these standard chips, each of which can take a researcher several days to characterize. The article says the group now regularly uses agents for routine measurements. In Yankelevich's words: "I can have agents running measurements for many hours overnight or while I'm working in the cleanroom. 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."^[S1]

The article also concedes that experienced researchers may still find the best calibration settings faster than current models; the gain is the time no longer spent monitoring every step, not a better calibration.^[S1]

Yankelevich says she has "built infrastructure to guide agents through several parts of my work - measurement, theory, and chip design - and now it's really starting to pay off." What does the work in this case is not only the model: measurement-specific skills, the chip's design targets and an interface that stores each result for the next step are what let a chain of interdependent measurements run unattended.^[S1]

For novel experiments she gives agents narrower goals and leans more on their ability to write, modify and test code for control, analysis and simulation; connecting agents directly to the lab lets the group revise code, test it against real measurements and complete longer stretches autonomously. The article shows no run for this part, so it is a description, not evidence.^[S1]

A standard calibration on clear signals is the easiest part to demonstrate; the hard part the article itself names (weak signals, noise, ambiguous physical results) is what decides how much supervision an agent can actually replace in a lab.

It did reach real hardware.

The colour maps and fits in the screenshot come from measured signals on one qubit of the chip, labelled q1: a Rabi chevron with a driven centre of 199.27 MHz, a 1D Rabi oscillation, a Ramsey chevron and its linear-branch fit. The bare frequency the model stored, 198.592407 MHz, agrees with the half-flux spectroscopy minimum of 198.598819 MHz to 6.4 kHz: a cross-check between two different kinds of measurement, and the closest this case comes to quantitative validation. But it is the difference between two measurements, not a calibration error against an independent true value. The screenshot gives no reference measurement, no repeat run and no uncertainty, so whether 6.4 kHz is good agreement cannot be judged from the public material. On this site's ladder that reaches rung 4: new measurements were executed on real hardware because the model chose them.^[S1]

But the evidence is a vendor case study.

The source is OpenAI's own site, and the screenshot is "an excerpted GPT-5.6 Sol chain-of-thought from a calibration run." What the screenshot measures is two steps on one qubit: a six-qubit chip does not mean six qubits calibrated, and 17 s and 67 s are the times of those two steps, not of the whole procedure. There is no paper, no code, no published skills and no statistics: how many chips, how many runs, how many of the six qubits were calibrated, how often a researcher stepped in, how much faster than a researcher: none of it has a number. "Little intervention on clear signals" and "the group now regularly uses agents for routine measurements" are statements by the article and the researcher, not measurements in the screenshot, which is why they sit under suggested.^[S1]

"Better" is not claimed.

The article says outright that experienced researchers may find the best settings faster. What is claimed is supervision time saved, not calibration quality beyond a person's.^[S1]

Interaction is the whole point of this case

Most AI-for-science claims stop at code and compute: the model writes programs, runs simulations, produces predictions. Here the model issued sweep ranges and pulse amplitudes to a real chip and read back signals from inside a dilution refrigerator. That is the distinction this site exists to make visible, even while the evidence is thin.

Execution is not the hard part; interpretation is

The boundary the article draws itself is weak, noisy, ambiguous signals. Standard calibration of a standard chip has already been written down as procedure, and the model does well inside it; once the physics does not show up as expected, a person is still needed.

The next evidence that would change the judgment

A record with statistics: how many chips, how many runs, how many researcher interventions, and time and outcome against manual characterization, or a novel experiment the agent designs and runs itself. Either would turn this from a case into a result that can be compared.