Which production server fits a low-traffic Flask app: Gunicorn with gthread, Uvicorn, or serverless?

I'm deploying a low-traffic Flask application (a few requests per minute at most) and want to minimize hosting costs without sacrificing correctness. The Flask documentation lists several deployment options but does not prescribe one, so the choice is left open.

The candidates I'm weighing are: Gunicorn with a single gthread worker, Uvicorn with async views on Flask 2.0+, and a serverless option such as AWS Lambda via Zappa. My endpoints are mostly I/O-bound (database calls and external APIs), and I rely on a few third-party extensions whose async compatibility I haven't audited. I also use before_request and teardown_request hooks, and I've read that these may behave differently under an ASGI lifespan model.

Constraints: I want to avoid the development server, keep memory footprint low, and would like to verify the decision with a benchmark rather than assumed behavior.

Specific questions:

• For a low-traffic, I/O-bound app, does a single Gunicorn gthread worker generally use less memory than one Uvicorn process, or does the difference only matter at scale?
• Which Flask extensions are known to have issues under async views, and is there a reliable way to check compatibility beyond reading each extension's documentation?
• For sporadic traffic, how should I weigh Lambda's cold-start latency against the cost savings of serverless, and are there documented limits (timeouts, payload size) that commonly break Flask assumptions?