> I collected a few different sets of statistics: static stats for the stdlib and for mypy, dynamic stats for running mypy and a few of the benchmarks in pyperformance.

I'm sure you've considered this, but I'd be *very* careful using opcode stats from the official benchmarks to inform work on these improvements. It's probably better to stick to in-sample data from the stdlib / mypy / etc. and use the benchmarks *only* for measuring out-of-sample results.

Otherwise we may just end up hacking the benchmark suite. :)