GPT-5.6 Sol in a quantum lab: a bounded test of agentic calibration

A September 2026 case study from OpenAI and MIT's Engineering Quantum Systems Group documents GPT-5.6 Sol, used through Codex, performing calibration work on a previously unmeasured six-qubit superconducting chip. The agent operated through the group's existing orchestration software, selecting measurement parameters, running hardware-facing routines, analyzing returned data, and deciding whether to refine a measurement or persist its result. This is concrete evidence of an agent participating in a live experimental workflow, not merely drafting code or summarizing laboratory records.

The reported result depends on substantial local infrastructure rather than a generic prompt. The technical paper says the researchers iterated for several months on experimental setup details, chip designs, measurement-specific skills, source-code access, and a Jupyter MCP connection to the in-house orchestration system. Those skills supplied execution templates, prerequisite calibrations, parameter-selection guidance, failure modes, and examples of successful and failed analyses. The case therefore illustrates a deployment pattern in which domain procedures and controlled tools carry much of the operational context around the model.

The relevant comparison is between roles in the calibration workflow, not between GPT-5.6 Sol and a named competing model. Standard qubit calibration consists of dependent measurements whose outputs set the next experiment; the paper notes that a new multi-qubit experiment can require months and thousands of preliminary measurements. In the group’s workflow, the orchestration system already runs instruments and extracts data, while the agent assumes portions of the human role of assessing results, choosing follow-up measurements, and updating settings. That division makes the result most applicable to software-mediated, repeatable experimental loops with observable intermediate outputs.

The limitations are as important as the automation result. The paper reports that the agent found all six resonators and that researchers intervened on four of 40 target measurements for four fixed-frequency qubits, but it also describes the test chip as relatively simple. OpenAI's accompanying account says weak or noisy signals took longer and sometimes required guidance from an experienced researcher. For technical users, the practical implication is not unattended general-purpose science: it is that agents may reduce monitoring work in tightly instrumented routines, provided teams preserve checkpoints, expert review, and a way to steer or halt experiments when the physical evidence is ambiguous.