A single session running for multiple hours is prohibitively expensive, as per Anthropic. Regardless of whether it just waits for a prompt or does something.
Is the high expense coming from cache misses? If their workload does need to wait for a long time before it can continue, I wonder would starting new sessions and having to re-read the contexts and results anyway be any cheaper.