GPT-5.6 Sol is being tested as a lab operator, not just a coding assistant

OpenAI says Beatriz Yankelevich, a graduate student in MIT's Engineering Quantum Systems Group, connected GPT-5.6 Sol through Codex to software that controls superconducting-qubit experiments. In the reported setup, the agent selected measurement parameters, ran measurements, analyzed the output, and either refined the next step or saved results for later stages of calibration. This is a concrete change from using a model solely to draft code or summarize lab notes: the system was connected to an existing experimental-control loop, although OpenAI's account is a case study rather than an independent benchmark of the model's laboratory performance.

The workflow matters because qubit characterization is sequential. OpenAI describes resonance-frequency measurement, pulse calibration, readout calibration, and coherence measurement as interdependent tasks, where one result shapes the next experiment. Yankelevich tested the system on an uncalibrated six-qubit chip and supplied measurement-specific Codex skills plus design targets. OpenAI reports that the agent completed standard measurement sequences with little intervention when the signals were clear, and that the group now uses agents for routine measurements. The evidence supports practical help with a bounded workflow, not a claim that the model independently designs or validates new quantum hardware.

The approach also fits a broader technical pattern rather than standing alone. A March 2026 preprint on large-language-model-assisted superconducting-qubit experiments describes an agent framework that uses procedural knowledge and instrument-control tools to automate resonator characterization and reproduce a quantum non-demolition characterization. Both accounts place the language model inside a constrained loop of documented procedures, software tools, stored data, and physical measurements. For model users, the important comparison is therefore with a scripted automation stack: GPT-5.6 Sol appears to add flexible sequencing and interpretation around defined skills, while the surrounding control software and experimental constraints remain essential.

The limitations are as important as the demonstration. OpenAI says GPT-5.6 Sol took longer to find suitable parameters on weak or noisy signals and sometimes required guidance from an experienced researcher; it also says experienced researchers may still choose the best calibration settings faster. The independent preprint similarly demonstrates specific characterization tasks rather than a general replacement for experimental judgment. Teams considering similar deployments should keep permissions narrow, preserve reviewable logs and stopping conditions, and treat the result as supervised automation for repeatable procedures rather than evidence of autonomous scientific discovery.