The subagent itself keeps its own model. Sessions with background subagents used about 5% less of the limit, so the same limit gets you about 6% more prompts.
It's live for most Pro, Max and Ultra users and for everyone on Team and Enterprise. Update to v2.1.316+.
