I'd take those benchmarks with a massive pinch of salt. 338µs per request for Flask is the bad case. 0.5ms of the request is spent in the framework, that's not even noise. The fastest framework I've used is one I wrote for coldfusion and that's the only time I've expected all my requests to be sub 100ms. And even then, the network latency turns that into noise.
Look at features before performance - both in the case of frameworks and version of Python to use (Python 3 is a lot nicer).