And they also tune the inference of the models behind it.

You're using different tooling every day. Hard to benchmark.