Three Models, One Pipeline: Claude Code's Per-Subagent Model Routing Cuts Costs 60-80%

Assigning Haiku 4.5, Sonnet 5, and Fable 5 to the right tasks in a multi-agent workflow cuts API costs 60-80% with no quality loss — here's the decision framework with copy-paste SDK code.

In Claude Code's Workflow SDK, every agent() call defaults to the session model (Sonnet 5). But Haiku 4.5 runs at $1/$5 per million tokens — half the cost of Sonnet 5 — and handles any task where reasoning depth isn't the bottleneck: file reads, searches, grep-equivalent scans, and simple classification. Reserving Fable 5 for decisions where a wrong answer is expensive keeps architectural reviews high-quality without applying that cost across every agent in the fan-out.

Decision framework:

Task	Model	Why
Search, grep, file reads	Haiku 4.5	Content retrieval — speed over depth
Implement, debug, write	Sonnet 5 (default)	Strong code reasoning at locked-in $2/$10
Architectural decisions	Fable 5	Complex judgment worth the extra cost
Binary checks (test pass/fail)	Haiku 4.5	One-token output; waste to use Sonnet

SDK code:

export const meta = {
  name: 'routed-workflow',
  description: 'Model-routed multi-agent example',
  phases: [{ title: 'Search' }, { title: 'Implement' }, { title: 'Review' }],
}

// Fan out cheap searches with Haiku
const findings = await parallel(files.map(f => () =>
  agent(`List all auth-related functions in ${f}`, {
    model: 'claude-haiku-4-5-20251001',
    label: `scan:${f}`,
    phase: 'Search',
  })
))

// Implement on default Sonnet 5
const impl = await agent(
  `Implement auth changes based on: ${findings.filter(Boolean).join('\n')}`,
  { phase: 'Implement' }
)

// Escalate one architectural review to Fable
const review = await agent(
  `Review this auth implementation for security regressions and propose alternatives:\n${impl}`,
  { model: 'claude-fable-5', phase: 'Review' }
)

Cost math example. 20 files × 3K tokens (Haiku) + 1 implementation at 50K tokens (Sonnet 5) + 1 architectural review at 25K tokens (Fable):

• Haiku searches: 60K tokens × $1/M = $0.06
• Sonnet implementation: 50K × $2/M = $0.10
• Fable review: 25K × ~$15/M = $0.38
• Total: $0.54 vs $1.50+ if everything ran on Fable

With Sonnet 5's pricing now locked in permanently at $2/$10, these economics are stable — build the routing now.

Sources: Claude Code model config — code.claude.com · Claude model pricing — anthropic.com · Multi-model routing guide — buildthisnow.com