If input tokens dominate the cost to that extent, this implies that major gains are possible by making better use of caching. You could basically ask the model to do a one-time "compaction" step including a dump of the relevant portions of the code, and use that as the cached prefix for a large amount of "swarm" subagent calls.