At least one instance is always running, regardless of the number of queued jobs.
The is a limitation of the AWS API. It can be worked around (e.g. cancelling and creating fleets as necessary); otherwise, this surprising limitation should at least be documented.
Is there a plan to implement fleet creation in AWS per need in the near future? Or it will just stay like this where it scoops reservation id and use at least one spot instance?
There is no need to pay for spot instance (even though they are 80% cheaper) when at least one of them stays idle all the time. 8 core, 15GB RAM is still around $70 per month.