What changed since March 2026. The March version of this page fitted three growth curves to 14 months of data and said the best fit was exponential (doubling every ~1.2–1.4 months) and that "the inflection point has not yet been observed." Five more months of data say otherwise: Claude Code commits went from in March to in — a doubling time of ~ months, not 1.2 — and month-on-month growth fell from +89% (March) to +% (). The March exponential predicted for . Sigmoid models now beat the exponential in fitting configurations; the fitted inflection sits at and the estimated ceiling for public Claude commits is per month (logistic–Gompertz range in the primary fit; wider across configurations, see the model table). This page also replaces the hand-collected Nov 2025–Mar 2026 numbers with a repeatable, windowed Search-API collector, fills the "missing" Oct 2025 month, adds GitHub's Copilot coding agent, and drops the developer count that could no longer be measured. Details in How we know.

Claude Code commits,

All tracked AI commits,

six-tool series; + Copilot agent shown separately

Doubling time, exponential fit

AI share of public commits,

of push events; ≈6–18% of commits depending on commits-per-push, see notes

The bend: Claude Code commits/month, actual vs. the March-2026 forecast vs. current fits (log scale; band = collector lo–hi; projection to )

The same curve on a linear axis: where on the S are we? (Claude Code; sigmoid fits to all data, projected to )

Month-on-month growth, Claude Code (from Aug 2025; Oct 2025 now filled)

Growth above +50%/month through March 2026; +14–26%/month since April. Search-API months carry ±10–15% measurement noise, so single-month bars are indicative, the level shift is not.

AI-attributed commits as a share of public push events

Denominator: monthly PushEvents from GH Archive (BigQuery). The commits-per-push ratio is genuinely uncertain: GitHub's Octoverse 2025 reports 986M commits for 2025 against ~1.0B push events in GH Archive (≈1:1), while the payload size field averages 3.8 (over-counting merges) and a March-2026 SemiAnalysis estimate that Claude Code was 4.5% of public commits implies ≈3 commits per push. Read the share as a range — ~6% of commits at 3:1, ~18% at 1:1 — until GitHub publishes a monthly commit denominator. The March page's "×2.5 code-volume adjustment" is dropped: agent commits are larger than human ones (Robbes et al. 2026), but no defensible multiplier exists for this denominator.

Commits by tool (log scale, not stacked; Copilot coding agent shown as two instruments — see notes)

Copilot coding agent is new on this page. It is visible two ways: commits authored by copilot-swe-agent[bot] on PR branches, and user-authored squash-merge commits carrying Co-authored-by: Copilot trailers (often 15–45 per merge). The two count different objects and are at most additive. Both peak in Mar–Apr 2026 (~1.2M combined) and fall ~3× by June; whether that is a usage decline or another authorship change is unresolved. Devin, Gemini Code Assist and Codex use different queries than the March page (the March values could not be reproduced), so those series step at Oct 2025.

Growth-model comparison (Claude Code; ordinary least squares in linear space, Feb 2025 → last full month, partial month excluded)

Fit	Model	AIC	R²	Doubling / ceiling	Inflection

Space	Start	n	Best by AIC	Exp. doubling (mo)	Logistic ceiling	Logistic t₀	Gompertz ceiling	Gompertz t₀

Reading the sensitivity table: linear-space fits (weighted toward the high-volume months, which is what a ceiling question needs) agree on an April 2026 inflection and a 22–30M ceiling regardless of start month. Log-space fits weight the tiny 2025 months equally and are unstable with 3-parameter sigmoids on ≤18 points — they are shown so nobody has to take our word for the primary choice. Empty ceiling = fit ran to the parameter bound (no ceiling found in that configuration).

How we know, and how well

Two instruments, spliced.

Jan–Sep 2025 — GH Archive via BigQuery. Exact commit-level counts: every PushEvent payload carried the commit list, so trailers (Co-Authored-By: Claude <noreply@anthropic.com>), body markers and author e-mails could be regex-matched. ~400 GB scanned per month.
Oct 2025 onward — GitHub commit Search API. In October 2025 GH Archive's PushEvent payload shrank to {repository_id, push_id, ref, head, before} — no commit list at all (payload volume: Sep 379 GB → Nov 96 GB → Jul 2026 27 GB). The March page called October a "data gap"; it was this format change, and the Search API puts October at ~954K Claude commits, on trend. From October on, nothing in GH Archive can identify a tool or count developers, which is why the "Developers using AI" card is gone: the last measurable value was 36,692 distinct public pushers in Aug 2025, and it would be misleading to keep showing it.

What the Search API's total_count actually is. Every heavy query returns incomplete_results: true — GitHub times out and returns an estimate. Measured properties: (a) the estimate is close to the sum of sub-windows when a window holds fewer than ~2–3M matches, and 30–40% low when it holds 15–20M (July 2026: whole-month calls 11.6M and 14.3M; four weekly windows summed to 18.6M); (b) the same query over the same window drifts ±10–15% between calls, occasionally 2×. The collector therefore splits each month into windows sized to hold <1.5M matches, queries each window three times, uses the median, and sums; lo/hi are the sums of per-window min/max. Bot-account queries (author:google-labs-jules[bot], author:copilot-swe-agent[bot], author:devin-ai-integration[bot]) return exact, repeatable counts. Every raw call is logged and published.

Cross-checks. Where both instruments overlap (Mar–Sep 2025) they agree to ±20%. The March-2026 hand-collected Claude values are reproduced by the collector (Feb 2026: 5.19M hand vs 5.44M; Mar 2026: 10.19M vs 10.29M). Independently, Quispe (Caltech, 2026) harvested 7,786,771 Claude co-authored commits for Jan 2025–Jan 2026 from GitHub's public API; this page's series sums to 7.83M over the same window.

What this measures and what it doesn't. Attribution markers, not AI use. Claude Code adds its trailer automatically (~80% of assisted commits carry it; Kraishan 2025), Copilot's inline completions add nothing (~9%), Cursor adds nothing. Cross-tool ratios here are largely disclosure-rate ratios. Public repositories only — GitHub reports 81.5% of contributions happen in private repos. Squash-merges collapse many agent commits into one. Khosravani & Mockus (2026) find single-channel detection undercounts Claude Code by ~30× at the project level. Everything on this page is a lower bound with a stated instrument, which is more than the "275 million AI commits a week" figures circulating without one.

Model fitting. Exponential, logistic and Gompertz curves fitted by nonlinear least squares (scipy.optimize.curve_fit) to Feb 2025 → last full month; Jan 2025 (24 commits, pre-launch) and the current partial month are excluded. AIC = n·ln(RSS/n) + 2k. The hindcast row refits the same models to data through Mar 2026 only, reproducing the March page's exponential (doubling 1.22 mo) and its "no ceiling found" logistic, and shows what those predicted for the months since.

References. Quispe, arXiv:2605.25438 · Khosravani & Mockus, arXiv:2606.24429 · Robbes et al., arXiv:2601.18341 · Kraishan, arXiv:2512.00867 · Ghaleb, arXiv:2601.17406 · GitHub Octoverse 2025. Code, raw Search-API calls, monthly CSV and this page: github.com/edgardomunoznajar/github_explosion. Machine-readable numbers: data.json.