The AI system achieved autonomous calibration and measurement workflows with humans intervening in only a few of 40 target measurements, demonstrating a highly automated, closed-loop quantum experimentation process.

OpenAI and MIT discuss broader implications, suggesting a shift toward AI-enabled ‘super AI + autonomous labs’ that reframes scientists’ roles toward high-level design and interpretation, while noting AI cannot fully replace expert intuition for highly ambiguous results.

The approach involves giving agents specialized lab capabilities to perform routine tasks efficiently.

MIT’s Engineering Quantum Systems Group uses agents for routine measurements, allowing researchers to focus on interpretation, experimental design, and planning next steps, increasing throughput.

Researchers provided the AI with comprehensive experimental context, chip designs, measurement templates, and failure modes, enabling it to drive from initial microwave pulse settings to resonance identification, Rabi calibration, Ramsey measurements, and coherence time estimates (T1 and T2).

Performance declined with weak, noisy, or ambiguous signals, occasionally requiring expert input, showing AI can automate structured experiments but struggles with uncertain or unfamiliar conditions.

OpenAI and MIT report that GPT-5.6 Sol, with Codex, autonomously conducts end-to-end quantum computing experiments on a six-qubit superconducting chip, including parameter inference, hardware control, data processing, and adaptive optimization.

GPT-5.6 Sol, via Codex, autonomously performed routine measurements on a MIT six-qubit superconducting chip, selecting settings, operating hardware, analyzing data, and deciding next steps, reducing need for constant supervision.

Experiments take place in a dilution refrigerator at millikelvin temperatures, with hardware control performed remotely via software, illustrating the real-world complexity of automating quantum lab operations.

Codex has specialized skill packs for specific measurement scenarios, with three AI roles—theory, control/monitoring, and simulation—operating in parallel to enable rapid, closed-loop experimentation.

The study highlights AI-driven science automation’s potential to shrink iteration cycles from weeks to hours or minutes, while noting current limitations in low signal-to-noise regions where human oversight remains necessary.

The AI agent achieved clear results by identifying qubit transition frequencies, calibrating control pulses, reading out qubit states, and measuring information retention times, with minimal guidance in clear-signal conditions.