Cost Breakdown

The Cost Breakdown panel shows how costs are distributed across your SDLC pipeline, helping you identify expensive stages and optimize resource allocation. It displays three key views: average cost per issue, cost per station, and human-vs-agent time distribution.

Cost Components

All cost calculations include:

• LLM provider costs - Claude, GPT, Gemini, and other AI models used
• Compute costs - sandbox execution (e2b, Docker, Vercel, Modal, etc.)
• Estimated human time - reviewer/approver hours (hourly rate configurable per org)

Cost is denominated in USD and is calculated at the session level when a work item completes, then aggregated by issue, station, and time period.

Left Panel: Cost Metrics

Avg Cost per Issue

The primary cost metric: total cost divided by number of completed issues.

Formula: Total Cost ÷ Total Completed Issues

Interpretation:

• <$0.50 - very efficient; mostly small tasks or cached LLM responses
• $0.50-$2.00 - typical range; healthy balance of automation and quality
• $2.00-$5.00 - higher complexity; multi-turn agent loops or expensive models
• >$5.00 - likely high-complexity work or expensive provider routing; review agent configuration

Example: 47 issues, $256.82 total cost → $5.47/issue average.

Cost per Station

A horizontal bar chart showing cumulative cost spent in each station.

Why stations vary in cost:

• Development - often most expensive; longest duration, most LLM usage
• Security - variable; may require expensive specialized models
• QA - depends on test automation; manual review is expensive
• Research - variable; depends on specification complexity
• Acceptance - typically cheap; short duration
• Deploy - typically cheap; automated

Reading the chart:

• Longer bars = more cost in that station
• Compare to throughput: if Development has highest cost but also highest throughput, it's efficient
• If a station has high cost AND low yield, it's inefficient (rework is expensive)

Cost Trend (Sparkline)

A small line chart showing daily cost over the time window.

Why it matters: Cost trend shows whether you're getting better (cost per issue declining) or worse (cost per issue rising).

Interpretation:

• Declining trend → agents improving, fewer loops, better reasoning
• Rising trend → degraded agent quality, more rework, or expensive model usage
• Flat trend → consistent cost; neither improving nor degrading

Right Panel: Time Distribution

Throughput

Items completed per day (efficiency metric).

Formula: Total Completed Issues ÷ Days in Time Window

Interpretation:

• Combine with cost: high throughput + low cost = highly efficient
• High throughput + high cost = fast but expensive
• Low throughput + high cost = slow and inefficient (needs attention)

Human vs Agent Time

A stacked bar or pie chart showing the split between human-authored and agent-authored work.

Human time:

• Manual issue triage and specification
• Code review and approval
• Manual testing
• Acceptance validation

Agent time:

• Issue analysis and requirements extraction
• Implementation (code writing)
• Security and QA automation
• Deployment automation

Interpretation:

• High agent % (>80%) - factory is highly automated; agents are doing most of the work
• Balanced split (40-60% agent) - hybrid workflow; good for high-trust environments
• Low agent % (<20%) - mostly manual; less automation benefit; consider increasing agent responsibility

Using Cost Breakdown

Identify Expensive Stages

Sort the "Cost per Station" bars mentally from longest to shortest:

1. Highest-cost station - Is this expected? Research and Development usually cost the most
2. Unexpectedly expensive station - If Security costs more than QA, investigate model routing
3. Low-cost stations - Are Acceptance and Deploy cheap? (They should be)

Cost per Issue Optimization

To reduce cost per issue:

Cost Efficiency Metrics

Calculate your own efficiency:

Cost Efficiency = Avg Yield ÷ Avg Cost per Issue

Example: 92% yield ÷ $1.23/issue = 74.8 points
Goal: >50 (high yield relative to cost)

Time Windows

All cost metrics aggregate over the selected time window (7d, 30d, 90d). Cost Breakdown metrics include:

• 7d - spot trends; noisy due to small sample size
• 30d - stable baseline; recommended for planning
• 90d - long-term trends; seasonal effects visible

Data Freshness

Cost metrics are updated hourly by the factory aggregation job. Cost data is calculated at session completion and should appear in the dashboard within 2-3 hours of issue completion.

Configuration

Cost attribution is configured per organization:

• LLM model pricing - from platform model catalog
• Compute pricing - per sandbox provider (e2b, Docker, Modal, Vercel)
• Human hourly rate - configurable per org or per role
• Attribution method - can weight stages by proportion of session duration

See Cost and Caps for budget and rate limits.

API Access

Fetch Cost Breakdown metrics programmatically:

curl -H "Authorization: Bearer $RENSEI_API_KEY" \
  "https://app.rensei.ai/api/factory/metrics?metricType=cost&timeRange=30d"

Response:

{
  "metrics": {
    "avgCostPerIssue": 1.23,
    "costPerStation": [
      { "station": "development", "cost": 0.62 },
      { "station": "qa", "cost": 0.38 },
      { "station": "security", "cost": 0.15 },
      ...
    ],
    "costTrend": [
      { "date": "2026-05-03", "cost": 1.18 },
      { "date": "2026-05-04", "cost": 1.25 },
      ...
    ],
    "humanTimePct": 25.0,
    "agentTimePct": 75.0
  }
}

See Metrics API for full details.

Cost Caveats

Next Steps

• To set per-(provider × authMode) cost caps, see Cost and Caps
• To route to cheaper models for specific task types, see Model Routing
• To track cost per issue over time, see Cost Per Issue