AI-generated code has crossed the majority threshold, and the engineers who stopped writing are the best ones

The number that gets quoted most is Shopify’s. Tobi Lütke told Harry Stebbings that the share of AI-generated code at Shopify is “a fair deal over 50%” and converting rapidly toward higher figures. That figure is striking on its own. What Lütke said next is more striking: many of the company’s best engineers have not written a single line of code since December. Not struggling engineers reassigned to other tasks. The best ones.

“Many of our best engineers have not written code this year. Ever since December. December changed everything. Opus changed everything.” Lütke did not specify a calendar year, and the claim should be read exactly as he made it: a personal and organizational inflection point tied to a month and a model, not a precise product-release date drawn from external record. The inflection is the point. The cause of it is secondary.

Anthropic’s own numbers are further along. Krishna Rao, speaking on Invest Like The Best, said that more than 90% of Anthropic’s code is now written by Claude Code. This is a company that builds the model doing the writing. The figure carries a particular weight because Anthropic has every incentive to understand what its model can and cannot do. When the people who built the tool report that it is writing nearly all of their own software, the claim is not promotional. It is operational.

Many of our best engineers have not written code this year. Ever since December. December changed everything. Opus changed everything. Tobi Lütke · 20VC with Harry Stebbings

The third data point in circulation is more complicated. Wesley Huff and Daniel Priestley, both guests on The Diary Of A CEO, each cited Spotify as a company whose engineers have stopped writing code since December. Both were repeating an assertion attributed to Spotify, not speaking as Spotify representatives. That claim has spread quickly on podcasts, but it is secondhand as the evidence stands: Spotify has not been heard from directly in this set of sources. The convergence in the telling is real. The originating confirmation is not yet on the record.

Even setting Spotify aside, the two confirmed data points from Lütke and Rao sketch the same shape. At Shopify, a commerce platform, the majority of code is AI-generated and the share is accelerating. At Anthropic, an AI lab, the share exceeds 90%. These are not comparable organizations, which is part of why the pattern across them matters. One builds merchant tools. The other builds frontier models. Both have crossed the threshold where human-written code is the minority contribution.

The more consequential observation is about which engineers have stepped back. In conventional software development, the engineers with the most experience tend to write the most code and guard the quality of what gets committed. Lütke’s framing inverts that assumption. The best engineers, by his account, are the ones who have moved furthest from the keyboard. That suggests the transition is not about replacing engineers who lacked the skill to keep up. It is about engineers with enough judgment to direct a model rather than type.

What that means for how engineering organizations are structured, hired for, and evaluated is a question the evidence does not yet answer. The percentages are a threshold, not a destination. Where the curve goes from a “fair deal over 50%” and from 90% is an open question, and none of the speakers claimed to know. What they did claim is that the transition has already happened faster than anticipated, that it was not gradual, and that a specific moment marked the before and after. Whether December was a month, a model, or both, the word keeps appearing across separate conversations on separate platforms. That pattern is the signal.