GPT-5.6 Sol Even Calibrates Qubits! Inside Quantum Computing Experiments with Codex

Hi everyone, I'm Shii-chan! Today I want to share a cool story straight out of a quantum computing lab!

OpenAI News openai.com

What was announced?

OpenAI's News shared how Beatriz Yankelevich, a graduate student in MIT's Engineering Quantum Systems Group (EQuS), used GPT-5.6 Sol together with Codex to see how far she could push automation in quantum computing experiments.

EQuS studies superconducting qubits, which run inside dilution refrigerators cooled to near absolute zero and are precisely controlled with microwave signals. Once a chip is fabricated, packaged, and cooled, researchers can only interact with it through software, which makes it a natural testbed for AI agents. By connecting Codex to the lab's experiment software, Yankelevich could hand off running measurements, analyzing the results, and deciding what to try next.

Why it matters

Preparing and running qubit experiments can take months, and just the preliminary measurements can number in the hundreds or thousands. Each measurement depends on the last, since one result shapes what gets measured next. Qubit properties can drift, and unexpected behavior can throw off results, so experienced researchers have always had to recognize these shifts and adapt on the fly. That mix of software control, repeated measurement, and adaptive decision-making is exactly why this looked like a compelling use case for AI agents.

What changes

Yankelevich tested GPT-5.6 Sol on an uncalibrated six-qubit chip, one of the standard types EQuS routinely fabricates to benchmark its process. When the signals were clean, Codex completed the standard measurement sequence with little intervention: identifying the qubit's transition frequencies, calibrating the control and readout pulses, and measuring how long it retained quantum information.

When signals were weak or noisy, though, GPT-5.6 Sol took longer to find suitable parameters and sometimes needed help from an experienced researcher. That suggests today's agents handle clearly defined experimental workflows well, but interpreting ambiguous physical results is still a challenge.

Here's how Yankelevich put it:

"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."

Dive Deep

Yankelevich gives Codex measurement-specific "skills" that explain how to run and evaluate each experiment. Combined with the chip's design targets, GPT-5.6 Sol uses these to choose measurement parameters, operate the hardware, analyze the resulting data, and decide whether to refine the measurement or move on to the next step.

Routine chip characterization follows a well-defined workflow, so it's easy to hand to agents. For novel experiments, she instead gives Codex narrower goals and leans more on its 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 get longer stretches of autonomous work done. Yankelevich says she's built infrastructure guiding agents through measurement, theory, and chip design work, and that it's now paying off: she can run multiple agents on different problems at once, while she spends most of her time on higher-level work like interpreting results, designing experiments, and planning the agents' next steps.

Wrap-up

• An MIT graduate student used GPT-5.6 Sol with Codex to run, analyze, and steer superconducting qubit experiments
• With clean signals, the agent handled routine calibration (transition frequencies, pulse calibration, coherence time) almost autonomously
• Weak or noisy signals were harder, sometimes requiring help from an experienced researcher
• Handing off routine chip characterization frees the researcher to focus on interpreting results, designing experiments, and planning next steps

Worth a read if you're curious about applying AI agents to a highly specialized experimental field like quantum computing, or about where agent autonomy currently hits its limits!