Can GPT-5.6 Sol calibrate a quantum chip with no human help?

Only when the signals are clean. OpenAI reports that GPT-5.6 Sol finished a standard measurement sequence with little researcher intervention on clear data, but took longer and sometimes needed an experienced researcher when signals were weak or noisy. OpenAI's own conclusion is that agents handle well-defined workflows while interpreting ambiguous physical results remains hard.

What did the agent actually measure on the six-qubit chip?

GPT-5.6 Sol identified each qubit's transition frequencies, calibrated the microwave pulses used to control and read the qubit, and determined how long the qubit held quantum information. For each step it chose the measurement parameters, operated the hardware, analyzed the returning data, and then either refined the measurement or saved the result to feed the next one.

How long does characterizing a superconducting qubit chip usually take?

OpenAI writes that preparing and running qubit experiments can take months and require hundreds to thousands of preliminary measurements, and that a single standard chip of the kind EQuS fabricates can occupy a researcher for several days. That repetition is what made calibration a candidate for agents in the first place.

Is the MIT quantum calibration work documented in more detail?

Yes. OpenAI links a technical case study, Agentic calibration of superconducting qubits, as a PDF alongside the blog post. The post itself also reproduces an excerpted GPT-5.6 Sol chain of thought from a calibration run and a plot of calibration measurements the model completed on its own, both credited to the EQuS group.

Does the MIT group still use agents after the experiment?

Yes. OpenAI reports that the EQuS group now regularly uses agents for routine measurements. Beatriz Yankelevich says she runs agents for many hours overnight or while she works in the cleanroom, checks on them from her phone, and steers them when something needs fixing or she wants to explore a different direction.

OpenAI · 2026-09-08 · major

GPT-5.6 Sol calibrates qubits — Codex agents run measurements at MIT

OpenAI published a case study in which GPT-5.6 Sol, driven through Codex and wired into lab software, ran the standard calibration sequence on an uncalibrated six-qubit superconducting chip in MIT's Engineering Quantum Systems Group.

OpenAI share card for the report on GPT-5.6 Sol running quantum computing experiments

An OpenAI case study puts a Codex agent on real quantum hardware and lets it calibrate a six-qubit chip on its own.

Quick facts

Publisher	OpenAI
Published	September 8, 2026
Model	GPT-5.6 Sol, harnessed to Codex
Lab	MIT Engineering Quantum Systems Group (EQuS)
Hardware	Uncalibrated six-qubit superconducting chip
Researcher	Beatriz Yankelevich, MIT graduate student
Known limit	Needs human help on weak or noisy signals

What is it?

Connecting Codex to the software that coordinates a superconducting-qubit experiment let GPT-5.6 Sol run measurements, read the results and decide what to try next. OpenAI published the write-up on September 8, based on work by Beatriz Yankelevich, a graduate student in MIT's Engineering Quantum Systems Group, on an uncalibrated six-qubit chip of the standard type the group uses to check its fabrication process.

How does it work?

Yankelevich gave Codex measurement-specific skills describing how to run and judge each experiment, plus the chip's design targets. From there the agent picked measurement parameters, drove the hardware, analyzed the returning signals and either refined the measurement or stored the result as an input to the next one. Because a packaged chip inside a dilution refrigerator is reachable only through software, the whole loop stayed inside the agent's reach.

Why does it matter?

Physics labs run on long chains of interdependent measurements where each result changes the next step, and that supervision is what eats a researcher's week. EQuS now leaves routine characterization to agents and puts people on experiment design and analysis. OpenAI is also plain about the ceiling: experienced researchers may still find the best calibration settings faster, and the agent could not interpret weak, noisy or physically ambiguous data without help.

Who is it for?

experimental physicists and researchers automating instrument control

Frequently asked questions

Can GPT-5.6 Sol calibrate a quantum chip with no human help?: Only when the signals are clean. OpenAI reports that GPT-5.6 Sol finished a standard measurement sequence with little researcher intervention on clear data, but took longer and sometimes needed an experienced researcher when signals were weak or noisy. OpenAI's own conclusion is that agents handle well-defined workflows while interpreting ambiguous physical results remains hard.
What did the agent actually measure on the six-qubit chip?: GPT-5.6 Sol identified each qubit's transition frequencies, calibrated the microwave pulses used to control and read the qubit, and determined how long the qubit held quantum information. For each step it chose the measurement parameters, operated the hardware, analyzed the returning data, and then either refined the measurement or saved the result to feed the next one.
How long does characterizing a superconducting qubit chip usually take?: OpenAI writes that preparing and running qubit experiments can take months and require hundreds to thousands of preliminary measurements, and that a single standard chip of the kind EQuS fabricates can occupy a researcher for several days. That repetition is what made calibration a candidate for agents in the first place.
Is the MIT quantum calibration work documented in more detail?: Yes. OpenAI links a technical case study, Agentic calibration of superconducting qubits, as a PDF alongside the blog post. The post itself also reproduces an excerpted GPT-5.6 Sol chain of thought from a calibration run and a plot of calibration measurements the model completed on its own, both credited to the EQuS group.
Does the MIT group still use agents after the experiment?: Yes. OpenAI reports that the EQuS group now regularly uses agents for routine measurements. Beatriz Yankelevich says she runs agents for many hours overnight or while she works in the cleanroom, checks on them from her phone, and steers them when something needs fixing or she wants to explore a different direction.

Try it

https://cdn.openai.com/pdf/case-study-agentic-calibration-of-superconducting-qubits.pdf