model-routing

Model Routing System

You have access to intelligent model routing. Before executing any task, analyze complexity and route to the appropriate model tier.

Current model IDs and pricing — Last verified: 2026-08-28. Re-verify at platform.claude.com before trusting these again; they have moved more than once in 2026. Haiku → claude-haiku-4-5 ($1.00 / $5.00 per 1M input/output) · Sonnet → claude-sonnet-5 ($2.00 / $10.00) · Opus → claude-opus-5 ($5.00 / $25.00, default). For the hardest long-horizon reasoning, escalate above Opus to claude-fable-5 ($10.00 / $50.00 — most capable; thinking is always on, no disabling it; reserve for genuinely demanding work). Model IDs are complete as written — never append a date suffix. The tiers below are model-agnostic.

Routing Decision Matrix

TASK COMPLEXITY ANALYSIS
────────────────────────────────────────────────────────────────

HAIKU (Fast, Cheap) - claude-haiku-4-5 - Use for: ├── Simple file operations (read, list, navigate) ├── Scaffolding and boilerplate generation ├── Deterministic transformations (format, lint, compile) ├── Status checks and health monitoring ├── SEO metadata generation ├── Deployment commands (after code is written) ├── Documentation formatting ├── Simple search and replace │ │ Token cost: $1.00/1M input, $5.00/1M output │ Latency: Fastest │ Use when: Task has clear, unambiguous steps

SONNET (Balanced) - claude-sonnet-5 - Use for: ├── Feature implementation (standard complexity) ├── Bug fixes requiring analysis ├── Content writing (articles, social posts) ├── Code review and quality checks ├── Test generation ├── Refactoring with clear patterns ├── API integration work ├── Database schema design │ │ Token cost: $2.00/1M input, $10.00/1M output │ Latency: Medium │ Use when: Task requires reasoning but not deep strategy

OPUS (Strategic, Complex) - claude-opus-5 - Use for: ├── Architecture decisions (system design) ├── Multi-agent coordination (council, swarm) ├── Strategic planning (business, product) ├── Complex debugging (multi-file, subtle bugs) ├── Security audits and vulnerability analysis ├── Enterprise AI system design ├── Book writing (narrative, character development) ├── Research synthesis (multiple sources) ├── Ambiguous requirements interpretation │ │ Token cost: $5.00/1M input, $25.00/1M output │ Latency: Slower; most capable of the routine tiers │ Use when: Task requires deep reasoning, creativity, or strategy

FABLE (Hardest, escalation only) - claude-fable-5 - Use for: ├── Long-horizon agentic work where Opus already fell short ├── The single hardest step in an otherwise-Opus-routed task │ │ Token cost: $10.00/1M input, $50.00/1M output │ Latency: Slowest; thinking is always on (cannot be disabled) │ Use when: Opus-tier reasoning wasn't enough — this is a ceiling, not a starting point

## Automatic Routing Rules

When processing a request, apply these rules:

### Route to HAIKU when:
- User says: "deploy", "format", "lint", "check status", "list", "scaffold"
- File patterns: `*.config.*`, `package.json`, `tsconfig.json`
- Commands: `/mcp-status`, `/inventory-status`, `/nextjs-deploy` (execution phase)

### Route to SONNET when:
- User says: "write", "implement", "fix", "create", "build", "test"
- File patterns: `*.ts`, `*.tsx`, `*.py`, `*.md` (content files)
- Commands: `/article-creator`, `/create-music`, `/spec`, `/generate-social`

### Route to OPUS when:
- User says: "design", "architect", "strategy", "council", "analyze", "research"
- Keywords: "enterprise", "system", "multi-agent", "complex", "strategic"
- Commands: `/starlight-architect`, `/council`, `/author-team`, `/research`

### Escalate to FABLE when:
- Opus already ran on the task and the result fell short
- The task is genuinely the hardest tier — long-horizon agentic work, not routine strategy
- Never route here first; Fable is a ceiling for escalation, not a default

## Cost Optimization

BEFORE (No routing): All tasks → Opus → $25.00/1M output tokens

AFTER (With routing): Simple tasks (40%) → Haiku → $5.00/1M = $2.00 Medium tasks (45%) → Sonnet → $10.00/1M = $4.50 Complex tasks (15%) → Opus → $25.00/1M = $3.75 ────────────────────────────────────────────── TOTAL: $10.25 vs $25.00 = 59% cost reduction

## Implementation in Task Tool

When using the Task tool, specify model based on routing:

```javascript
// Simple task - use haiku
Task({
  subagent_type: "Explore",
  model: "haiku",
  prompt: "List all files in src/"
})

// Medium task - use sonnet (default)
Task({
  subagent_type: "code-reviewer",
  model: "sonnet",
  prompt: "Review this PR for issues"
})

// Complex task - use opus
Task({
  subagent_type: "Plan",
  model: "opus",
  prompt: "Design the architecture for a multi-tenant SaaS platform"
})

Command-Level Routing

Command	Default Model	Rationale
/acos	sonnet	Router needs reasoning
/article-creator	sonnet	Content creation
/create-music	sonnet	Creative work
/infogenius	sonnet	Research + creation
/starlight-architect	opus	Strategic design
/council	opus	Multi-perspective
/research	sonnet	Information synthesis
/spec	sonnet	Feature planning
/nextjs-deploy	haiku	Execution
/mcp-status	haiku	Status check
/inventory-status	haiku	Status check
/publish	haiku	Execution
/polish-content	sonnet	Editing
/review-content	sonnet	Quality check

Escalation Pattern

If a haiku-routed task fails or produces poor results:

1. Automatically escalate to sonnet
2. If still failing, escalate to opus
3. If opus itself falls short on a genuinely hard task, escalate to Fable — last resort, not a retry habit
4. Log escalation for learning

haiku (attempt) → fail → sonnet (retry) → fail → opus (retry) → fail (rare) → fable (last resort)

Relationship to other routing configs

This skill is the estate's model-facts reference — current model IDs, context windows, and per-token pricing. It is not the only routing config in the estate, and on task-class routing decisions it does not have the final word:

• starlight-evals/routing-table.json is the evidence-derived source of truth for task-class routing — which tier a task class should hit, backed by eval-round evidence and confidence scores rather than the judgment calls in this file. Where its route for a task class disagrees with the "Route to X when" rules above, the table wins.
• agentic-ops-hub/fleet/model-routing.json and a referenced ~/.starlight/routing.toml are operational routing configs elsewhere in the estate. They route by abstract tier name (opus, sonnet, ...) and should resolve those names against this skill (or the evals table) rather than hardcoding their own model IDs or pricing.
• This skill's job stays narrow: keep the model IDs, context windows, and pricing above correct and current. It is not merged with the other three configs here — this section exists only so the next agent knows which source to trust when they disagree.

---

Model Routing v1.1 - Implementing claude-flow's intelligent routing pattern. Last verified: 2026-08-28.