Claude Code's changelog: fixes now outnumber features

A developer graphed every entry on the Claude Code changelog page and counted 385 versions and 5,132 bullet points between 0.2.21 on April 2, 2025 and 2.1.260 on September 3, 2026. The analysis, published by Bonsai, classifies each line by the verb it starts with.

At a glance

• The method is simple: every bullet on the changelog page was sorted by its opening verb, with added, fixed and improved as the three named buckets and everything else left unclassified.
• April 2025 produced 56 bullets across 27 versions, while August 2026 produced 735 across 28, the same release cadence carrying roughly thirteen times as much content inside it.
• Releases landed on 258 days across the eighteen months, with 27 of those falling on a Saturday and only 6 on a Sunday, a pattern the author reads as unusual consistency.

Claude Code is closed source, so its changelog is the only outside measurement of how the team spends its time, and a verb histogram is a crude instrument. Still, the shape it produces reads as a product past its expansion phase: the surface is largely built, and the work has moved to keeping it standing. The growth in volume without a matching growth in release count also fits a workflow where drafting is cheap.

The verb mix shifted over the period. In July 2025, 41% of bullets started with "added". By mid 2026 that share sits between 6% and 10%, while the proportion beginning with "fixed" has not dropped below 56% since February.

The author notes that Claude Code is built with Claude Code, and attributes a good chunk of the output to agents doing the work with a human pressing the button. Because the product is closed source, the changelog page is the only visibility outsiders have into how the team operates.

Two caveats come with the data. April 2, 2025 shows 17 releases on a single day, which the author describes as the changelog being seeded rather than a real day of work. The "Other" bucket is not a category but the classifier giving up on lines that do not start with added, fixed or improved.

What the changelog cannot show

The unclassified lines are the obvious place to look for human authorship, and the author would like to claim them as such but cannot verify it from outside the company. The dataset stops at version 2.1.260 on September 3, 2026, and nothing in it separates machine-written entries from hand-written ones. That distinction stays out of reach while the repository remains closed.