Cost Analysis

Process Steps

1. Usage data collection -- Gather token usage records from episode memory, execution logs, or provided usage data. For each invocation, record: model tier, input tokens, output tokens, mode, task type, and timestamp.

2. Per-model cost calculation -- Apply pricing to each invocation:

   • Input cost = input_tokens * input_price_per_token
   • Output cost = output_tokens * output_price_per_token
   • Total cost = input_cost + output_cost
   • Aggregate by model tier, mode, and task type

3. Cost distribution analysis -- Identify where costs concentrate:

   • Which model tier accounts for the most cost?
   • Which mode (build, research, review) is most expensive?
   • Which task types have the highest per-task cost?
   • What percentage of cost is input tokens vs output tokens?

4. Efficiency metrics -- Calculate efficiency ratios:

   • Cost per successful task completion
   • Cost per line of code produced (for build tasks)
   • Cost per review (for inspector tasks)
   • Retry cost (wasted tokens on failed attempts)
   • Token waste ratio (tokens spent on tasks that were later replanned)

5. Tier optimization analysis -- For each task type, evaluate if a cheaper model tier could handle it:

   • Research tasks: could use efficient tier instead of capable?
   • Simple build tasks: could reduce thinking budget?
   • Standard reviews: could use lighter review mode?
   • Estimate cost savings for each optimization

6. Thinking budget analysis -- Analyze thinking budget usage:

   • Distribution of thinking modes (low/medium/high/xhigh) by task type
   • Correlation between thinking budget and task success rate
   • Identify tasks where high thinking was used but low would suffice
   • Identify tasks where low thinking led to failures (under-invested)

7. Trend projection -- Based on historical data, project future costs:

   • Daily/weekly/monthly cost trends
   • Growth rate of token usage
   • Projected cost for the next period
   • When the budget will be exhausted at current rate

8. Optimization recommendations -- Produce actionable recommendations:

   • Model tier downgrade opportunities with risk assessment
   • Thinking budget adjustments with expected savings
   • Task batching opportunities (reduce per-invocation overhead)
   • Caching opportunities (avoid re-computing similar tasks)

9. Report assembly -- Compile the cost analysis report with: summary metrics, per-model breakdown, optimization opportunities, trend projections, and recommendations.

Output Format

The skill produces a cost analysis report:

{
  "success": true,
  "output": {
    "summary": {
      "total_cost_usd": 12.45,
      "period": "2025-01-01 to 2025-01-31",
      "total_invocations": 342,
      "avg_cost_per_task": 0.036
    },
    "by_model": {
      "claude-opus": {"invocations": 45, "cost": 8.20, "pct": 65.9},
      "claude-sonnet": {"invocations": 180, "cost": 3.50, "pct": 28.1},
      "claude-haiku": {"invocations": 117, "cost": 0.75, "pct": 6.0}
    },
    "by_mode": {
      "build": {"invocations": 89, "cost": 5.60},
      "code_discovery": {"invocations": 120, "cost": 2.10},
      "code_review": {"invocations": 65, "cost": 3.80},
      "plan": {"invocations": 68, "cost": 0.95}
    },
    "optimization_opportunities": [
      {
        "opportunity": "Route code_discovery tasks to haiku tier",
        "current_cost": 2.10,
        "projected_cost": 0.45,
        "savings_usd": 1.65,
        "risk": "low -- discovery tasks are pattern-matching, not reasoning-intensive"
      }
    ],
    "projection": {
      "next_month_estimate": 14.20,
      "growth_rate": "14% month-over-month",
      "budget_exhaustion": "Not at risk (budget: $50/month)"
    }
  },
  "metadata": {
    "data_source": "episode_memory",
    "records_analyzed": 342
  }
}

Quality Standards

This skill is governed by the following standards from the skill registry:

• STD-WRK-018: Cost analysis MUST include per-model token breakdown and total estimated cost
• STD-UNI-001: All skill outputs MUST conform to the declared output_schema
• STD-UNI-003: All skill executions MUST log entry and exit at INFO level with timing
• CON-WRK-008: Cost analysis must not access external billing APIs without explicit permission
• CON-WRK-009: Operations modes run post-execution -- MUST NOT modify already-reviewed code

Examples

Example: Monthly cost review with optimization

Input: