Yankelevich Shows GPT 5.6 Sol Can Run Superconducting Qubit Measurements Without Constant Supervision
MIT graduate student Beatriz Yankelevich used OpenAI's GPT 5.6 Sol with Codex to autonomously run superconducting qubit calibration workflows, a public case study that lab leaders can use to scope agentic automation pilots.
Classy aggregates publicly available material; we did not conduct a private interview.
What changed
On 8 September 2026, OpenAI published a case study from MIT's Engineering Quantum Systems Group showing graduate student Beatriz Yankelevich using GPT 5.6 Sol connected to Codex to run superconducting qubit measurement workflows with limited human supervision. The setup linked Codex to existing lab control software so the model could execute measurements, read outputs, and choose follow up steps inside the same experimental loop.
Yankelevich works on superconducting qubits cooled in dilution refrigerators and controlled through microwave signals. OpenAI reports that once the chip is packaged and cooled, experiments are software driven, which made the workflow a natural test for an agent that can call tools rather than only draft text.
Why it matters
Quantum labs spend months on calibration sequences that repeat across chips and cooldown cycles. If a frontier model can reliably execute routine measurement branches, principal investigators can shift senior researcher time from button pushing to experiment design and failure analysis. That is a staffing and throughput decision, not a chatbot upgrade.
The case also shows where agentic AI is heading in other instrument heavy domains: closed loops where each step depends on the last reading, with human review at defined checkpoints rather than continuous babysitting.
Who is affected
University quantum hardware groups, national lab qubit teams, and vendors selling control electronics or cloud access to early quantum processors should treat this as an operations template. Venture and corporate R&D leaders funding physical AI should compare Yankelevich's workflow against their own lab software APIs before buying generic copilots.
What to do next
Audit one recurring calibration script in your lab and list which steps already expose a programmatic interface. Pilot agent access only on that branch with logging, rollback, and a hard stop before any parameter change that affects cryogenic safety or high power microwave limits.
What to watch
Whether OpenAI, IBM, Google Quantum AI, or IonQ publish independent replication studies with error rates and wall clock savings on different qubit modalities beyond Yankelevich's superconducting setup.
Sources
- Primary. OpenAI, How GPT 5.6 Sol helps run quantum computing experiments (8 September 2026). Describes Yankelevich's MIT workflow, Codex integration, and autonomous measurement runs.
- Secondary. The Stack Observer, OpenAI's Quantum Lab Agent Shows Where Agentic AI Is Heading (9 September 2026). Places the case study in the broader agentic operations pattern.