Used hardware energy counters (CPU RAPL + GPU NVML) with WattSeal for per-process attribution rather than CPU% proxies. The Google Meet result was unexpected enough that we re-ran it in multiple browsers to make sure it wasn't a Firefox artifact.
Here are the results: https://wattseal.com/blog/power-benchmarks/video-call-power-comparison/#results
Raw data and benchmark procedures: https://github.com/Daminoup88/WattSeal-benchmark