Aider: The AI Pair Programmer Writing 88% of Its Own Code

Most "built with AI" stories are about a human using a chatbot to ship a product faster. Aider is the more unsettling version: an open-source AI pair programmer whose own README now claims that 88% of the code in its latest release was written by Aider itself — tracked, badge and all, under a metric the project calls "Singularity."

There's no company behind it in any visible sense. The GitHub organization that owns the repo, Aider-AI, lists no public members. Every recent tagged release is attributed to a single account, paul-gauthier. This is what a nearly one-person AI tool looks like when it's asked to keep building itself — including the parts that are starting to creak.

What Aider actually is

Aider is a terminal-based AI pair-programming tool: you point it at a local git repository, describe a change in plain English, and it edits files, runs the diff through git, and writes the commit message — one commit per change, so every AI edit is individually reversible. Per its own README, it works across roughly 100 programming languages and is model-agnostic, with recent commits adding support for models like Claude Opus 4.7 and GPT‑5.5 as they've shipped. It's free and Apache-2.0 licensed — the tool itself costs nothing, but you bring your own API key (or an OpenRouter account) and pay the model provider directly for tokens.

That "no license fee, pay for tokens" model is why the README's other badge matters: Aider says it processes roughly 15 billion tokens per week across its user base, and its PyPI package has logged about 6.8 million installs. Both numbers are self-reported, pulled straight from the project's own badge set rather than an audited dashboard — treat them as the maintainer's account of usage, not a third-party count.

One name on the releases

The most recent tagged release, v0.86.0, was published by a single account: @paul-gauthier. On PyPI, the package's listed maintainer is also paul-gauthier, and the project's homepage link on PyPI points straight back to the GitHub repo. His GitHub profile shows a Pro account, 20 repositories, and 1.6k followers — nothing that reads as a funded startup with a comms team.

None of that proves Aider is a solo operation with zero outside help — a repo with 4.9k forks and 476 open pull requests clearly draws outside contributors, whether or not most of those PRs end up merged. But the public-facing ownership, release cadence, and package metadata all point to one person as the person actually shipping it, which is a genuinely different shape than the venture-backed AI coding tools it competes with.

The "Singularity" metric, and how it's actually measured

The 88% figure isn't a marketing claim pulled from thin air — the project has published its methodology. A 2024 blog post in the repo lays out how it works: Aider runs git blame across its own codebase and counts how many lines trace back to a commit that Aider itself authored. At the time of that post, the number was small — "about 7% of the code now in the repo is attributable to an aider commit using git blame," out of roughly 14,000 lines and 4,000 commits. The post is upfront about the method's limits, too, noting that routine reformatting by the black code formatter "is likely obscuring aider's authorship of many lines" — which would make the true number higher than the git-blame count shows.

That post has since been updated with a header note that the 7% figure is outdated and the real number had already climbed to 70%. The README's current badge puts it at 88% for the latest release specifically — not the whole repo's history, a distinction worth holding onto, since "88% of this release" and "88% of everything ever committed" are different claims and the badge is measuring the former.

Whatever the exact number, the trend line — 7% in mid-2024, 70% not long after, 88% now — is a real, if self-reported, measure of a tool eating more and more of its own development over about two years. It's the same idea behind shipping a first AI agent workflow and then measuring how much of the follow-up work the agent actually handles, just applied recursively to the tool doing the automating.

The part that doesn't fit the highlight reel

Two things in the primary sources don't match a clean "one dev, infinite leverage" story.

First, Aider's GitHub releases page has been stuck on v0.86.0 since August 9, 2025 — over a year before this piece was written. But the PyPI package page shows a newer version, 0.86.2, published February 12, 2026, and commit history on the main branch shows ordinary development activity — new model support, repo-mapping fixes — as recently as May 2026. Read together, that looks like GitHub's tagged-release process fell behind the actual PyPI publishing cadence rather than the project going quiet, but it's a real gap between the two record-keeping systems, and it's the kind of thing that should make a prospective user check the commit log, not just the "Releases" tab, before assuming a tool is stale.

Second, the repo currently carries 1.3k open issues and 476 open pull requests. For a project whose visible ownership is one GitHub account, that's a large backlog, and it's a fair question whether "88% of new code written by the AI" is solving the bottleneck that actually matters here — human review and triage capacity — or just moving where the bottleneck sits.

What this generalizes to

The transferable idea for other solo builders isn't "let the AI write most of your code" — plenty of people already do that with Claude Code, Cursor, and the rest of the growing field of AI agent tooling without publishing a number for it. It's that Aider turned "how much is the AI actually doing" into a metric with a published method, run against its own history, that anyone can go check with git blame. Most AI-coding claims — "80% AI-written," "built in a weekend" — arrive with no way to verify them. Aider's number can be wrong in the details (the post itself says so), but it's at least falsifiable, which is more than most vibe-coding revenue screenshots can say. If you're going to claim AI wrote most of your product, the credible version of that claim comes with a method attached, not just a percentage.