Something to consider having mapped all of these discussions on my one use-case.
Sub-interpreters allow some limited form of parallelism.
I see from The updated PEP 703 discussion that many people have shared their experience and metrics with their library/application running with no-gil. I didn’t see as much (or any?) people sharing similar experience/metrics having done so with sub-interpreters. I suspect that’s because they don’t’ yet support rich use-cases, like sharing data (I remember from the PyCon talk that plans are in the works for that to be the case).
I wonder if those rich-use cases were supported (ideally soon) if people would be happy with the experience/performance of using sub-interpreters. Then the discussion becomes slightly richer, as we might see that Option 1 feels less like a tradeoff because sub-interpreters provide real benefit for parallel use-cases.