Does Kaggle bill idle minutes on paid kernels, and can a low-traffic workload avoid them?

A recurring notebook workload on Kaggle runs only a few short jobs per day, with long idle gaps between executions. The goal is to keep total compute cost as low as possible while avoiding a full environment rebuild on every run.

Two platform behaviors shape the decision. Each kernel execution starts an isolated Docker container, and compute resources are not retained across separate runs, so state has to be re-initialized every time. Paid compute is also billed per minute of allocated instance time, which leaves it unclear whether idle minutes on a running session accrue charges. Free-tier sessions have historically shut down after roughly 90 minutes of inactivity, though current limits may differ from that baseline.

Does Kaggle bill idle minutes on a paid kernel session, or only while code is executing? Is the inactivity timeout enforced on paid tiers, or can a session persist between runs? For a few short runs per day, is re-initializing a fresh kernel each time cheaper than keeping one session alive?