OpenAI GPT-5.6 Sol enables autonomous quantum computing experiments

TL;DR

OpenAI released GPT‑5.6 Sol, a Codex‑powered AI agent that can autonomously execute, analyze, and iterate on superconducting qubit measurements, cutting down weeks‑long calibration work to unattended overnight runs.

Autonomous measurement workflow

GPT‑5.6 Sol was connected to the MIT Engineering Quantum Systems (EQuS) lab software stack, giving the model direct control over microwave pulse generation, data acquisition, and analysis. The agent received a library of measurement‑specific skills, selected appropriate parameters, operated the dilution‑refrigerator‑cooled chip, and processed the resulting signals without human intervention for routine tasks.

"I can have agents running measurements for many hours overnight or while I’m working in the cleanroom," said graduate student Beatriz Yankelevich.

Calibration of a six‑qubit chip

When presented with an uncalibrated six‑qubit superconducting chip, GPT‑5.6 Sol:

  1. Identified transition frequencies from clear signal data.
  2. Calibrated control and readout pulses (Rabi and Ramsey sequences).
  3. Measured coherence times (T₁, T₂) and stored results for subsequent steps. The agent completed the full calibration sequence with minimal researcher input when signal‑to‑noise ratios were high.

Limitations with noisy data

In cases of weak or noisy measurements, GPT‑5.6 Sol required more iterations to locate suitable parameters and occasionally needed expert guidance. This demonstrates that current AI agents excel at well‑defined workflows but still struggle with ambiguous physical signals.

Impact on researcher productivity

EQuS typically spends several days characterizing each standard chip. By delegating routine measurements to GPT‑5.6 Sol, researchers now:

  • Run experiments unattended for extended periods.
  • Monitor progress remotely via a phone interface.
  • Reallocate time to higher‑level tasks such as data interpretation, experiment design, and theory development.

"I spend most of my time on higher‑level work—interpreting results, devising experiments, planning next steps for the agents," Yankelevich noted.

Extending the agent’s role

Beyond routine calibration, Yankelevich assigns narrower experimental goals to Codex agents, leveraging their code‑generation capabilities to:

  • Write and modify control scripts.
  • Implement custom analysis pipelines.
  • Simulate theoretical models against live data. Multiple agents can operate concurrently on distinct problems, further amplifying throughput.

Implications for quantum research

The demonstration shows that large language model agents can bridge the software‑hardware gap in quantum labs, turning lengthy, repetitive calibration into a largely automated process. While expert intuition remains valuable for noisy or novel experiments, the time savings and scalability offered by GPT‑5.6 Sol could accelerate the iteration cycle for superconducting qubit development and, by extension, the broader quantum computing roadmap.

Image credits: EQuS group

Sources