Cost Tracking Framework

When This Activates

This skill activates when:

• User asks about API costs or spending
• Concerns about expensive operations
• Need to optimize token usage

Token Cost Reference

Claude Pricing (Approximate)

Model	Input (per 1M tokens)	Output (per 1M tokens)
Opus	~$15	~$75
Sonnet	~$3	~$15
Haiku	~$0.25	~$1.25

Typical Operation Costs

Operation	Tokens	Approximate Cost
Simple question	500-2K	$0.01-0.05
File read + analysis	2-10K	$0.05-0.25
Code generation	5-20K	$0.15-0.50
Multi-file refactor	20-100K	$0.50-2.50
Long conversation	50-200K	$1.00-5.00

Cost Optimization Strategies

1. Route to Local LLM (FREE)

Use local_ask for simple tasks:

# FREE - no API cost
local_ask question="where is the login function?"
local_ask question="explain this error" mode=explain
local_review file="src/auth.ts" focus=bugs

Good for local:

• Simple lookups ("where is X?")
• Code explanations
• Commit message generation
• Quick code reviews

2. Use Memory Tools First

Pre-indexed memory is instant and free:

# Instant, no API cost
memory_query "authentication flow"
memory_functions name="handleLogin"
smart_read path="src/auth.ts" detail=summary

3. Reduce Context Size

• Use smart_read with detail=summary before detail=full
• Truncate large files to relevant sections
• Clear conversation when changing topics

4. Batch Related Questions

Instead of 5 separate messages, combine:

"Can you: 1) explain the auth flow, 2) find the login
component, 3) check for security issues, and 4) suggest
improvements?"

Gateway Metrics

Check current efficiency:

gateway_metrics format=summary

Returns:

• Cache hit rate
• Token savings
• Routing breakdown (local vs API)

Cost Estimation

Before expensive operations:

This refactor will touch ~20 files.
Estimated cost: $0.50-1.00
Proceed? [Y/n]

Budget Awareness

Daily Patterns

• Morning: Fresh context, lower cost
• Long sessions: Context grows, higher cost
• After compaction: Reset context, lower cost

High-Cost Triggers

• "Analyze entire codebase"
• "Review all files in directory"
• "Generate comprehensive documentation"
• Very long conversations (>50 turns)

Saving Tips

1. Start fresh for new topics - Don't carry irrelevant context
2. Use subagents - They have focused context
3. Check memory first - Summaries save full file reads
4. Compress transcripts - Archived sessions are compressed
5. Local for simple tasks - Ollama is free

Monitoring Commands

# Check gateway efficiency
python3 ~/.claude-dash/learning/efficiency_tracker.py --report

# View session sizes
du -sh ~/.claude-dash/sessions/*