Live dashboard: botcommits.dev
Claude Code went from 24 publicly-attributed commits in January 2025 to over 5.2 million per month in February 2026 — a 216,000x increase in 13 months. This project tracks that growth across four AI coding tools using real commit-level data.
| Metric | Value |
|---|---|
| Claude commits, Jan 2025 | 24 |
| Claude commits, Feb 2026 | 5,187,311 |
| Growth multiple | 216,000x in 13 months |
| Claude market share | 96% of detectable AI commits |
| Total AI commits tracked | ~13M (cumulative) |
| Best fit growth model | Exponential (R²=0.989, AIC=314.7) |
| Logistic ceiling found? | No — converged to upper bound (~1B) |
| Peak repos with AI commits | 54,291 (Aug 2025) |
| Peak developers using AI | 36,692 (Aug 2025) |
- Claude Code — Co-Authored-By trailers, commit body markers, author email (
noreply@anthropic.com) - Aider —
aider:prefix in commit messages, Co-Authored-By trailers - Devin —
devin-ai-integrationactor (BigQuery),co-authored-by devin-ai(Search API) - OpenAI Codex CLI — Co-Authored-By trailers, author email (
noreply@openai.com)
GitHub Copilot and Cursor do not leave commit-level attribution markers and are invisible to this methodology. The numbers here are a lower bound.
- Jan–Oct 2025: GH Archive via Google BigQuery — exact commit-level counts from PushEvent payloads (~400 GB/month scans)
- Nov 2025–present: GitHub Search API — GH Archive stopped including commit message payloads in Nov 2025, reducing tables from ~400 GB to ~89 GB
- Total push events: BigQuery aggregation (no payload scan, cheap)
github_explosion/
├── dashboard/
│ ├── index.html # Live dashboard (Chart.js, single-page static)
│ ├── Dockerfile # nginx:alpine container for Cloud Run
│ └── favicon.png # Custom bot icon
├── analysis/
│ ├── claude_commits.py # BigQuery query runner (--query, --months flags)
│ └── fit_models.py # Exponential / Logistic / Gompertz curve fitting (scipy)
├── queries/
│ └── claude_commits.sql # 6 BigQuery queries (timeseries, top repos, model breakdown, all tools, etc.)
├── data/
│ ├── claude_all_tools.csv # AI commits by tool (Jan–Oct 2025, 4 tools × 10 months)
│ ├── claude_total_pushes.csv # Total push events per month (2024–2026)
│ ├── claude_timeseries.csv # Claude-only monthly timeseries
│ ├── claude_top_repos.csv # Top repositories by Claude commit count
│ └── claude_sample.csv # Sample commits with full metadata
├── requirements.txt
└── README.md
Three models fitted to Claude commit counts using scipy.optimize.curve_fit:
| Model | R² | AIC | Parameters |
|---|---|---|---|
| Exponential y = ae^(bt) | 0.989 | 314.7 | a=7506, b=0.502 |
| Logistic y = L/(1+e^(-k(t-t₀))) | 0.989 | 316.7 | L=1B (hit upper bound), k=0.504, t₀=23.5 |
| Gompertz y = Le^(-e^(-k(t-t₀))) | 0.981 | 323.8 | L=641M, k=0.093, t₀=30.0 |
The exponential model wins by AIC. The logistic model's ceiling parameter converged to its upper bound, meaning the data shows no evidence of saturation yet. This does not predict indefinite exponential growth — it means the inflection point of the S-curve has not yet been observed with 14 months of data.
# Install dependencies
pip install -r requirements.txt
# Run a BigQuery query (requires gcloud auth + project with BigQuery access)
python -m analysis.claude_commits --query all_tools --months 202501
# Dry run (shows query + estimated bytes, no charge)
python -m analysis.claude_commits --dry-run --query all_tools --months 202501
# Fit growth models to existing data
python -m analysis.fit_models
# Run the dashboard locally
cd dashboard && python3 -m http.server 9443- First load (14 months, all tools): ~$30 in BigQuery scan costs
- Monthly update (1 month, all tools): ~$2–3 per month
- Total push counts: ~$0.10 per query (no payload scan)
- Dashboard hosting: Cloud Run free tier
- Oct 2025: GH Archive data appears incomplete (61M push events vs ~71M adjacent months). Excluded from model fitting and growth rate calculations.
- Nov 2025+: GitHub Search API returns approximate
total_countvalues, not exact counts. - Private repos: Only public repos are measurable. 81.5% of GitHub activity is in private repos (GitHub Octoverse 2025). The dashboard shows a speculative upper bound extrapolation.
- Attribution bias: Claude Code automatically adds Co-Authored-By trailers. A paper (arXiv:2512.00867) found 80.5% of Claude-assisted commits include attribution vs 9% for Copilot. This methodology inherently favors tools with automatic attribution.
Contributions welcome — open an issue or submit a PR. Areas where help is especially useful:
- Additional AI tool detection signals
- Historical backfill to 2024
- Copilot detection via behavioral fingerprinting
- International/regional breakdowns
- Academic analysis and peer review
Open source. Data, queries, and code are free to use for research, journalism, and analysis. Attribution appreciated.