claude-code-strategy

Key Limitation: No Effort/Thinking Parameter Access

Critical finding: Claude Code custom commands and Task subagents support model selection but NOT effort/thinking parameters.

What This Means

• Opus 4.5 in Claude Code runs at default (high) effort
• Haiku 4.5 in Claude Code runs with thinking OFF (default)
• Cannot access the 39% cost savings from effort: medium
• Cannot enable Haiku extended thinking for complex tasks

Claude Code-Specific Pricing

Since we can’t use Opus medium effort, the cost comparison changes:

Key insight: Opus high = Sonnet cost, but Opus has better quality!

Revised Strategy for Claude Code

┌─────────────────────────────────────────────────────────────┐
│  MAIN CONVERSATION: Opus 4.5 (high effort - default)        │
│  Or: Sonnet 4.5 for cost-conscious users                    │
└─────────────────────────────────────────────────────────────┘
                           │
                    Task Subagents
                           │
       ┌───────────────────┴───────────────────┐
       ▼                                       ▼
┌─────────────────┐                 ┌─────────────────┐
│   HAIKU 4.5     │                 │   OPUS 4.5      │
│   (Simple)      │                 │   (Complex)     │
├─────────────────┤                 ├─────────────────┤
│ Cost: $6/M      │                 │ Cost: $18/M     │
│ Quality: ~70%   │                 │ Quality: 80.9%  │
│ Speed: Fastest  │                 │ Speed: Fast     │
├─────────────────┤                 ├─────────────────┤
│ • Install deps  │                 │ • Schema design │
│ • Run tests     │                 │ • Refactoring   │
│ • File ops      │                 │ • Architecture  │
│ • Verification  │                 │ • Integration   │
│ • Simple docs   │                 │ • Critical code │
└─────────────────┘                 └─────────────────┘

Why Sonnet is Now Obsolete in Claude Code

Since Opus high effort = Sonnet cost ($18/M):

Verdict: For complex tasks in Claude Code, always use Opus 4.5 - same cost, better quality.

Custom Command Configuration

Simple Task Command (Haiku)

---
model: claude-haiku-4-5-20251015
description: Run tests and report results
allowed-tools: Bash(bun test:*), Read
---

Run the test suite and provide a summary of results.

Complex Task Command (Opus)

---
model: claude-opus-4-5-20251101
description: Architectural analysis and refactoring
allowed-tools: Read, Write, Edit, Glob, Grep
---

Analyze the codebase architecture and propose improvements.

Planning Command (Opus)

---
model: claude-opus-4-5-20251101
description: Create detailed implementation plan
allowed-tools: Read, Glob, Grep, Write
---

Create a comprehensive plan for the requested feature.

Task Subagent Model Selection

In the Task tool, specify the model:

// Simple task → Haiku
Task(
  subagent_type: "general-purpose",
  model: "haiku",
  prompt: "Install dependencies and verify setup"
)

// Complex task → Opus (NOT Sonnet)
Task(
  subagent_type: "general-purpose",
  model: "opus",
  prompt: "Design the Zod schema with .passthrough() for flexibility"
)

Cost Comparison: Claude Code vs Direct API

For Complex Tasks (18 tasks like frontmatter-improvement)

Claude Code (no effort parameter):

Planning: 1 × Opus high = $18/M
Simple (11 tasks): 11 × Haiku = $66/M
Complex (7 tasks): 7 × Opus high = $126/M
Total: $210/M

Direct API (with effort parameter):

Planning: 1 × Opus high = $18/M
Simple (11 tasks): 11 × Haiku = $66/M
Complex (7 tasks): 7 × Opus MEDIUM = $77/M
Total: $161/M

Savings: 23% cheaper with direct API

When to Use Direct API Instead

Consider direct API calls when:

1. Running high-volume batch operations
2. Need the 39% cost savings from Opus medium
3. Building custom tooling outside Claude Code
4. Cost optimization is critical

Updated Execution Pattern

For the /execute Slash Command

---
model: claude-opus-4-5-20251101
description: Execute a plan with optimal model routing
allowed-tools: Task, Read, Write, Bash, Glob, Grep
---

Execute the plan at .claude/plans/$ARGUMENTS.md

Task routing:
- Simple/mechanical tasks → model: haiku
- Complex reasoning tasks → model: opus
- All tasks → NO Sonnet (Opus same cost, better quality)

Example Task Classification

Speed Considerations

Haiku advantages:

• 3-5x faster than Opus for simple tasks
• Lower latency for quick operations
• Ideal for parallel execution of many simple tasks

Opus advantages:

• Better first-attempt success (fewer retries)
• Token efficient (generates less verbose output)
• Better for sequential complex reasoning

Monitoring and Optimization

Track These Metrics

1. Task success rate by model

   • Haiku failures on “simple” tasks → reclassify as complex
   • Opus overkill on tasks → downgrade to Haiku

2. Cost per successful task

   • Haiku: Should be ~$0.06 per task
   • Opus: Should be ~$0.18 per task
   • Watch for retry overhead

3. Total execution time

   • Haiku parallelization benefits
   • Opus single-attempt efficiency

Adjustment Triggers

• Haiku failure rate >10%: Reclassify tasks or use Opus
• Opus tasks too simple: Downgrade to Haiku
• Cost exceeds budget: Review task classification

Summary: Claude Code Optimal Strategy

Future Considerations

If Claude Code Adds Effort Parameter Support

The strategy would change to:

• Simple tasks: Haiku (unchanged)
• Complex tasks: Opus medium (39% cheaper than current)
• Critical tasks: Opus high (unchanged)

Potential savings: ~23% reduction in complex task costs

Feature Request

Consider requesting Anthropic add effort/thinking parameters to:

• Custom slash command YAML frontmatter
• Task subagent configuration
• Claude Code settings

Last updated: 2025-11-24 Key insight: In Claude Code, Opus 4.5 (high effort) costs the same as Sonnet 4.5 but delivers +3.7pp better quality. Always prefer Opus over Sonnet for complex tasks.