Data · measured 10 Nov 2026

How much reverse engineering code do AI models write?

190,658 of the 450,474 commits in 838 public reverse engineering projects credit an AI, or 42% of them. Claude is credited on 183,907 and GPT / Codex on 10,022.

We count the AI co-author credits that coding tools write into each project's public git history. Every listed project page shows its own figures, and the census is rerun about once a week.

Commits checked
450,474
Commits crediting an AI
190,658
Projects with at least one
691
Share from the 10 busiest projects
39%
By model

Which models the commits credit

ModelCommitsProjects
Claude Opus 5.5106,567509
Claude (version not given)29,962164
Claude Sonnet 5.513,126126
Claude Opus 510,333123
Codex9,71334
Claude Opus 4.89,511136
Claude Opus 4.78,75452
Claude Fable 5.14,28093
Claude Opus 4.64,17051
Claude Fable 53,06380
Claude Opus 4.52,3917
GitHub Copilot1,63933
Claude Sonnet 51,51547
Cursor1,11021
Claude Sonnet 4.686637

A commit crediting two models counts once for each. These figures show how much a model is used, and the models page covers how well they do on benchmarks.

By week

AI-credited commits week by week

Reports from 27 October 2026 said Anthropic had steered Claude Code away from recreating specific existing games (PCWorld). In the week of 19 Oct 2026, 31,889 commits credited an AI across 271 projects. In the week of 2 Nov 2026, the latest full week, it was 27,827 commits across 332 projects. These totals cover every kind of project, games included.

Week startingAI-credited commitsActive projectsFirst AI commit that weekClaudeGPT / Codex
2 Nov 202627,82733211326,8221,974
26 Oct 202628,1382939427,7292,319
19 Oct 202631,88927110230,8872,386
12 Oct 202626,60420210625,1102,256
5 Oct 202616,4071064116,121633
28 Sept 20266,91968166,715148
21 Sept 20264,43654114,29029
14 Sept 20262,4173851,9735
7 Sept 20262,0216591,808125
31 Aug 20262,3885092,30376
Method

What this can and can't tell you

  • It is a minimum. A commit counts only if it names an AI as co-author or author, and many tools and people leave that out.
  • Commit counts measure activity. An agent that commits every few minutes counts for far more than a person who squashes a week of work, and the 10 busiest projects account for 39% of AI-credited commits.
  • It doesn't show whether the result works. For that, see each project's reported outcome.
  • Each commit counts once. Forks repeat their parent's history, so commits are matched by hash and forks are grouped into one project by their first commit.
  • The repositories are the 859 that our discovery had found on GitHub by 10 Nov 2026: 508 listed in the directory and the rest awaiting review. Recently active repositories are easier to find, so recent weeks can look busier than they were. Figures for individual projects are published only for listed entries.
Reuse

Download and cite the data

The full snapshot, including figures for each listed project, is one JSON file: download the data. It is free to reuse under CC BY 4.0 if you credit Reversed with AI and link to this page.

Reversed with AI, "AI commit census for reverse engineering projects", measured 10 Nov 2026, https://reversedwithai.com/ai-commit-census