Benchmarks of FastAPI vs async Flask?

I'm a developer without an interest in benchmarking and I'm trying to decide whether I should use Flask or FastAPI to build some Python/Vue projects. I'm seeing stuff online about how FastAPI was faster than Flask because Flask was single-threaded or something like that, whereas FastAPI was async, but apparently more-recently Flask added async routes, and so now I'm wondering if FastAPI is still(?) faster than Flask.

Has anyone done benchmarking tests comparing FastAPI to Flask async routes? I can't find any when I search Google.

According to a benchmark study by Miguel Grinberg,
FastAPI can be faster or slower than async Flask, depending on the web server and the Flask async type.
Generally Flask on a Greenlet powered WSGI server (Meinheld / Gevent) can offer comparable throughput as an async-first ASGI framework like FastAPI.
Note that Grinberg is comparing the overall performance of three parts: the Framework, the Web server and the Web application.

Here are the relevant frameworks' throughput in the three different scenarios he tried (a higher number is better):

* Meinheld is a WSGI server written in C.

Grinberg points out in his article that the exact relative results between frameworks will depend on the particular load that the server is put under, but his takeaway is that there isn't a large difference between frameworks:

> remember that the difference in performance between different frameworks or web servers isn't going to be very significant, so choose the tools that make you more productive!

More about Async approaches of Flask

Flask adopts different approaches to async view.
In Grinberg's benchmark tests, three approaches have been implemented: Aioflask (standard python ASGI with uvicorn), Greenlet in Meinheld and Gevent. Since the Aioflask test was not fully finished, only the result of Test 1 is available, which agrees with the Flask documentation:

> Flask’s async/coroutine support is less performant than async-first frameworks due to the way it is implemented.

Be aware that the Greenlet approach requires a special way of writing asynchronous operations, instead of standard python coroutines. Therefore, existing asyncio codebase needs to be patched with adapters like greenletio.
Gevent relies heavily on proper monkey patching the python standard libraries, database engines and other performance critical libraries.

Grinberg addressed multiple times this in his blog:
> Gevent tests did reasonably well in my benchmark and terribly in the original one. This is because the author forgot to patch the psycopg2 package so that it becomes non-blocking under greenlets.