Model Usage Tracking

Track LLM usage, token consumption, and estimate costs.

Monitoring

Token Usage

Track tokens used in conversations:

• •Input tokens
• •Output tokens
• •Total tokens

Cost Estimation

Anthropic Pricing (as of 2026)

• •Claude Opus 4.5: $15 per MTok input, $75 per MTok output
• •Claude Sonnet 3.5: $3 per MTok input, $15 per MTok output

OpenAI Pricing

• •GPT-4 Turbo: $10 per MTok input, $30 per MTok output
• •GPT-4: $30 per MTok input, $60 per MTok output

Usage Examples

User: "How many tokens have I used today?"

1. •Query session history
2. •Sum token counts
3. •Present totals

User: "Estimate my costs for this month"

1. •Gather usage statistics
2. •Calculate costs per model
3. •Present breakdown

User: "Which model is most efficient for my use case?"

1. •Compare token usage patterns
2. •Calculate cost per conversation
3. •Recommend most efficient model

Best Practices

• •Log token usage with each API call
• •Store usage data in session metadata
• •Provide cost estimates before expensive operations
• •Suggest model switching for cost optimization

Data Structure

usage = {
    "model": "claude-opus-4-5",
    "input_tokens": 1500,
    "output_tokens": 500,
    "cost_usd": 0.06,  # Estimated
    "timestamp": "2026-01-27T00:00:00Z"
}

Reporting

Generate usage reports:

• •Daily/weekly/monthly summaries
• •Per-model breakdown
• •Cost trends
• •Token efficiency metrics