Pick a Subagent's Model by Measuring It on Your Own Tasks
The slug says run cheap subagents and reserve Opus for the hard part. That is a reasonable hypothesis, not a rule. Finding the file that defines a config key and explaining why a loss curve diverges at step 3,000 both start with reading code; only the first is easy. Choosing a model per task by measurement is the kind of cost judgment an applied-AI team expects you to make.
Custom Claude Code subagents can set a model. The subagent documentation says the model field accepts sonnet, opus, haiku, fable, a full model ID, or inherit, and that Claude Code resolves the model from the per-invocation parameter first, then the subagent's frontmatter, then the CLAUDE_CODE_SUBAGENT_MODEL environment variable. It also says the built-in Explore agent inherits the main conversation's model, capped at Opus on the Claude API, and that a project subagent named Explore with model: haiku keeps exploration on a lower-cost model.
The protocol. Define two copies of one read-only subagent that differ only in model. Write 30 questions about your repository with answers you have checked, mixing lookups ("where is the tokenizer loaded?") with reasoning ("why does eval accuracy differ between batch sizes 1 and 8?"). Run both on all 30. Record correct or not, tokens and wall-clock time.
The comparison. The runs share questions, so compare pair by pair: count questions only the first model got right and questions only the second did. For each model's rate, the 95% interval is 1.96 × √(p(1 − p)/30), which at p = 0.8 is ±0.14, so small differences at n = 30 will not be conclusive. Split the results by lookup and reasoning questions; the answer may differ between them.
The cost. Price the accepted answer, not the call: expected cost = cost per attempt + (1 − s) × R, where s is the success rate and R is what it costs you or a stronger model to repair a wrong answer. A call that costs a third as much but succeeds half as often can cost more per accepted answer once R is large. The book's delegation chapter derives when a second agent pays at all.
This site has not run this comparison, so it reports no winner.
