The encoder benchmark
Codec comparisons usually quote somebody else's hardware. This one uses yours: the same sixty frames pushed through every encoder your browser exposes, timed, with the output sizes measured. The gaps you see are the difference between silicon and software on this machine.
What this machine can encode, and how fast
60 identical frames · 480p · 1,200 kbps target · your hardware
keep this tab in front while it runs
What is actually being measured
Every codec gets identical work: the same sixty decoded frames at the same resolution, aiming at the same bitrate. Only the encoder changes. That makes the frames-per-second column a fair comparison of how much work each codec demands, and the size column a rough read on what that work buys.
It is a measurement of your machine, not of the codecs in general. A laptop with an AV1 encode block will post numbers a desktop without one cannot approach, and the same browser on the same silicon can choose hardware for one codec and software for another. That is the point: the answer to "should I encode AV1 here" depends on the hardware in front of you, and this measures it rather than guessing.
Reading the gaps
A codec running many times slower than its neighbours is almost certainly running in software. Newer codecs demand more computation by design, since finding better compression means searching harder, so AV1 in software against H.264 in hardware is the widest gap you are likely to see. When two codecs land close together, both are probably being handled by dedicated blocks.
Output size tells the other half. A codec that is slower but meaningfully smaller is doing exactly what it promised; one that is slower and no smaller is being asked for a bitrate it did not need. The AV1 vs H.264 comparison takes that trade apart properly with quality metrics on a fixed corpus, which is the part a browser cannot honestly measure.
Why the numbers move between runs
Encoders share the machine. Thermal limits, other tabs, whether the laptop is on battery, and whatever else wants the GPU all move these figures, sometimes by a lot. Run it twice before believing any single number, and treat the ordering as more reliable than the absolute values.