I think the use case where the per-config-request pricing model could break is with serverless (like Lambda or Cloud Functions or ...). Secret management there is super important, and your solution looks great, except that what makes these platforms great for low-cost deployments is that they "scale to 0" (turn off your instances when there's no traffic), which means that each instance potentially lives just briefly. Since each start of each instance is a config request, that could in some cases become problematic.

That said, at your lowest tier you could have ~20 moments of "booting a new instance" per hour. That's probably plenty for most cases, although you may find users for whom it isn't enough.