Skip to content

4.1 Environmental Costof the AI‐Assisted Development ClioDeck

Frédéric Clavert edited this page Jul 24, 2026 · 5 revisions

NOTE: This estimation has been made by querying Claude to estimate its own environmental cost. It is probable that it is underestimated.

⚠️ Unit inconsistency in the methodology — genuinely unresolved, and the evidence points in two different directions. Both "Method 1" and "Method 2" below state their rate as gCO2e per 1,000 tokens, then label the arithmetic result of that rate kg CO2e with no conversion shown — e.g. Method 1: 2.34M tokens × 0.0001–0.0003 gCO2e/1,000 tokens works out to 0.234–0.702 grams, not the "0.234-0.702 kg CO2e" the page states.

Two pieces of evidence disagree about which side of this is the mistake:

  • The page's own downstream reasoning (totals, driving-distance equivalences) is built consistently around the kg-scale figures, and "6–11 km of driving" is a plausible order of magnitude for a car (~150–250 g CO2/km) — suggesting kg was the intended scale.
  • Published estimates of LLM inference carbon cost, however, are typically reported in grams or milligrams, not kilograms — e.g. a full ChatGPT request is commonly cited around 4 g CO2, and per-token costs in the single-digit-milligram range. Under that convention, the page's literal gCO2e rate is the plausible one, and it's the downstream kg-scale totals and equivalences that would be the actual error — inflated by roughly 1000× rather than the rate being 1000× too small.

No specific citable source for this page's exact rate was found either way, so this cannot be settled from the page's own content. Treat every number below as uncertain by a factor that could run either direction — on top of the staleness noted next.

⚠️ Historical snapshot — not current. Every figure below rests on the repo's size as of January 16, 2026: 104 commits, ~19,742 lines in src/. As of this correction, the repository has 292 commits and ~69,923 lines in src/*.ts(x) alone — roughly 3× the commits and 3.5× the code volume this whole estimate is built on. Since every downstream number (tokens, kg CO2e, driving-distance equivalents) is derived multiplicatively from those two inputs, the totals below are stale by a comparable factor, not a rounding error. This page has not been recomputed since; treat it as a record of the estimate's method, not of the project's current footprint.

Project Overview

  • Project: Cliodeck
  • Development period: December 19, 2025 - January 16, 2026 (~4 weeks)
  • Total commits: 104
  • Lines of code (src directory): ~19,742 lines
  • Total project files: ~61,512 lines (including config, dependencies metadata, etc.)
  • Development approach: AI-assisted development with Claude Code

Estimation Methods

Method 1: Conservative Estimation Based on Commit Activity

This method estimates usage based on the number of commits and typical AI interaction patterns.

Assumptions:

  • Average prompts per commit: 5 (accounts for feature development, debugging, iterations)
  • Average tokens per exchange: ~4,500 tokens (1,500 input + 3,000 output)

Calculations:

  • Total commits: 104
  • Estimated prompts: 104 × 5 = 520 prompts
  • Total estimated tokens: 520 × 4,500 = ~2.34 million tokens

Carbon footprint:

  • Claude Sonnet: 0.0001-0.0003 gCO2e per 1,000 tokens
  • Estimated emissions: 0.234-0.702 kg CO2e
  • Average: ~0.47 kg CO2e

Method 2: Code Volume Analysis

This method estimates usage based on the volume of code produced and the typical read/write ratios in AI-assisted development.

Assumptions:

  • AI reads 3-5x more code than it writes (for context understanding)
  • Code requires 2-3 iterations per feature on average
  • Lines of code generated: ~20,000 lines

Calculations:

  • Code reading/context: 20,000 × 5 = 100,000 lines read
  • Code generation with iterations: 20,000 × 2.5 = 50,000 lines generated
  • Estimated total tokens: ~5-10 million tokens

Carbon footprint:

  • Claude Sonnet: 0.0001-0.0003 gCO2e per 1,000 tokens
  • Estimated emissions: 0.5-3 kg CO2e
  • Average: ~1.75 kg CO2e

Comparison Summary

Method Estimated Tokens CO2 Emissions (kg) Equivalent
Method 1: Commit-based 2.34M 0.47 kg CO2e ~2 km of driving
Method 2: Code volume 5-10M 1.75 kg CO2e ~7 km of driving

Best estimate range: 0.47-1.75 kg CO2e

Additional Environmental Costs

Computer Usage

Development on 3 different computers over ~4 weeks:

Assumptions:

  • 3 computers
  • ~50-100 hours of development time per computer
  • Average power consumption: 50-100W per computer

Calculations:

  • Total energy: 3 × 75 hours × 75W = 16.875 kWh
  • Carbon intensity (France): ~60 gCO2e/kWh
  • Computer usage emissions: ~1.0 kg CO2e

Total Project Environmental Cost

Component CO2 Emissions (kg)
AI assistance (conservative) 0.47 kg
AI assistance (code volume) 1.75 kg
Computer usage 1.0 kg
Total (conservative) 1.47 kg CO2e
Total (higher estimate) 2.75 kg CO2e

Context and Comparisons

Equivalencies

The total environmental cost of developing Cliodeck (~1.5-2.75 kg CO2e) is approximately equivalent to:

  • 6-11 km of driving a standard car
  • 1-2 days of average laptop use (8 hours/day)
  • 0.5-1 kg of beef production
  • 1-2 hours of streaming video in HD

Alternative Development Approach

Traditional development without AI assistance:

Estimated time: ~160 hours of development (4 weeks × 40 hours/week)

Energy consumption:

  • 160 hours × 75W = 12 kWh
  • Carbon emissions: 12 kWh × 60 gCO2e/kWh = ~0.72 kg CO2e (computer only)

Comparison:

  • AI-assisted development: 1.47-2.75 kg CO2e
  • Traditional development (computer only): 0.72 kg CO2e
  • Additional AI cost: +0.75-2.03 kg CO2e

However, this doesn't account for:

  • Potentially longer development time without AI assistance
  • Additional computer usage from longer development cycles
  • Energy costs of research, documentation reading, etc.

Methodology Limitations

Challenges in Accurate Estimation

  1. Conversation history loss: Repository name change resulted in lost conversation history
  2. Multiple computers: Usage distributed across 3 different machines
  3. No direct token tracking: Claude Code CLI doesn't provide built-in usage metrics
  4. Context windows: Actual token usage may be higher due to large context windows
  5. Revision iterations: Difficult to estimate exact number of code revisions

Improving Future Estimates

For more accurate tracking in future projects:

  1. Use Anthropic API directly: Provides detailed token usage metrics
  2. Manual session logging: Keep a development journal with session times
  3. Token tracking tools: Implement custom usage tracking
  4. Single repository: Maintain conversation history in one location

Research Sources

Environmental impact estimates based on:

  • Claude model specifications (Anthropic, 2025)
  • Research on LLM carbon footprints (2024-2025)
  • Grid carbon intensity data (France: ~60 gCO2e/kWh)
  • Standard computing power consumption metrics

Conclusion

The environmental cost of AI-assisted development of Cliodeck is estimated at 1.5-2.75 kg CO2e, with the AI assistance component contributing approximately 0.5-1.75 kg CO2e. This represents a modest environmental footprint, comparable to a short car journey.

While AI assistance adds some environmental cost compared to traditional development, the productivity gains and potential reduction in development time may offset this through reduced overall computer usage time. The total impact remains relatively small in absolute terms.


Analysis date: January 16, 2026 Project: Cliodeck v1.0 Analysis based on git history and estimated usage patterns

Clone this wiki locally