Entire

Mostly comes down to cost vs. capability per task type:

Why Sonnet is the sensible default for subagents:

~5x cheaper than Opus, and subagent costs compound fast when you spawn several in parallel
Most subagent work is bounded: grep/search, file reads, focused research, narrow implementation — Sonnet 4.6 handles these fine
Faster turnaround, which matters because the parent is blocked waiting on foreground agents
The subagent's job is often to filter raw output into a summary; the parent (Opus) does the synthesis

When Opus subagents earn their cost:

Open-ended investigation where the subagent has to make judgment calls (e.g. "is this migration safe?")
Code review / security review on non-trivial diffs
Debugging that requires hypothesis generation across many files
Anything where a wrong summary back to the parent silently corrupts your decision — Sonnet hallucinating a "looks fine" defeats the point

The asymmetry that makes the default reasonable: if Sonnet gets a search task wrong, you notice quickly and re-run. If Opus gets a search task right, you paid 5x for the same answer. Reserve Opus for tasks where you can't easily verify the output.

Practical heuristic: lookup/search/mechanical → Sonnet (or even Haiku). Judgment/synthesis/review → Opus, especially when the parent will act on the result without re-reading source.