Edge Copy setup link

Subagent Model Picker

Give each Claude Code subagent the cheapest model that still gets its task right: Haiku, Sonnet, Opus or Fable, with a one-line reason per task.

Works in

What your agent can get

Claude Code subagents
A tier per task
Cheapest model that passes
Checkable results first
One escalation, then take it back
Aliases pinned to model IDs

How it works

Step 1
Sort

Asks three questions about each task.

Step 2
Pick

Maps the tier to Haiku, Sonnet, Opus or Fable.

Step 3
Spawn

Passes that model on every subagent.

Example

Complete the four tasks in TASKS.md. Delegate each task to its own subagent and use the subagent-model-picker skill to choose each subagent's model.

✓ Used Edge: Subagent Model Picker

Inventory and config on Haiku, the date bug on Sonnet, the allocation on Opus

lead claude-opus-5-5 loads subagent-model-picker, then spawns 4 subagents

haikuInventory undocumented functions$0.050
sonnetFix parse_date bug$0.086
opusImplement money.allocate$0.207
haikuConvert config.ini to JSON$0.050

35/35 hidden checks, visible tests green

Tally: two tasks on Haiku, one on Sonnet, one on Opus, with no retries needed.

$0.716 for the run, lead $0.324 included · 28% less on average over 15 runs

Recorded runs, 14 Oct 2026: the median-cost run of each arm, 15 runs per arm.

Evidence

Measured: 28% cheaper, same pass rate

15 runs with the skill against 15 with every subagent on Opus: $0.73 against $1.01 per run on average, and all 30 runs passed 35 of 35 hidden checks. Read the full eval

Give your agent Subagent Model Picker.

Give each Claude Code subagent the cheapest model that still gets its task right: Haiku, Sonnet, Opus or Fable, with a one-line reason per task.

Copy setup link