90% of Anthropic Code Is Written by Claude

The Developer Who Stopped Writing Code for Two Months

Boris Cherny is the person in charge of Claude Code at Anthropic. He leads their AI coding tool. And what he posted on X shook the industry.

"I have not written code in over two months."

100% of his code is written by Claude Code and Opus 4.5. He submitted 22 PRs yesterday and 27 the day before. All of it was 100% Claude-generated code. The person building the AI coding tool does not write code himself. That is the reality of February 2026.

An Anthropic spokesperson disclosed the company-wide numbers. Between 70% and 90% of code is AI-generated. The Claude Code project itself is roughly 90% written by Claude Code. AI is building itself.

Microsoft CEO Satya Nadella reported 30% back in April 2025. In just 10 months, Anthropic has tripled that figure. The pace is terrifying.

Spotify Developers Are Not Writing Code Either

This is not an Anthropic-only story. On February 12, 2026, Spotify co-CEO Gustav Soderstrom made a stunning statement during an earnings call.

"Our best developers have not written a single line of code since December."

Spotify runs an internal system called Honk. It is a real-time remote code deployment system built on Claude Code. Here is how it works.

A developer commutes to work and sends a Slack message: "Fix this bug in the iOS app." When Claude finishes, the completed app build is sent back via Slack. The developer reviews it and merges to production. Features are deployed before they even arrive at the office.

Using this approach, Spotify shipped over 50 new features in 2025. AI-powered Prompted Playlists, Page Match for audiobooks, and About This Song were all built this way.

This aligns with Anthropic releasing Claude Opus 4.5 in December 2025. Opus 4.5 crossed the critical threshold for coding capability.

So What Do Developers Actually Do

If they are not writing code, what do developers do all day? Boris Cherny's workflow tells the story.

The core is design and review. Before any code is generated, Cherny meticulously plans with AI. He asks "Is this really the best approach?", pushes back on the AI's answer, and iterates. Only when the plan is sufficiently refined does he instruct execution.

Execution is done by AI. Cherny reviews the output. If something is wrong, he instructs again. This loop repeats.

Gergely Orosz of The Pragmatic Engineer summarized it well.

"The developer's role is shifting from 'code writer' to 'AI orchestrator'."

The Dark Side of 10x Productivity

Does everything get better when speed increases? Steve Yegge's warning is worth paying attention to.

Yegge is a veteran engineer who spent decades at Amazon and Google. He recently published a piece on Medium called "The AI Vampire".

"AI excites you and lets you capture enormous value. Then at some point, you fall asleep."

Yegge himself experiences what he calls "nap attacks" — suddenly passing out during the day. His friends, startup founders, report the same thing. Exhaustion hits after AI coding sessions.

He compares the phenomenon to slot machines. AI delivers a constant stream of dopamine. Code is done — excitement. Bug is fixed — excitement. Feature is added — excitement. The brain cannot keep up with this pace.

Yegge's advice is clear.

"Vibe coding tops out at 3 hours a day. Anything beyond that is unsustainable."

Companies need to recognize this. A 10x productivity gain can lead to 10x workload demands. But the human brain does not scale 10x.

The AI Coding Paradox

There is a fascinating study. It comes from METR (Model Evaluation and Threat Research), published in January 2026.

When experienced developers used AI tools, tasks actually took 19% longer. Yet they believed they were 20% faster.

Why the disconnect?

The volume of AI-generated code is massive. But the time spent reading, understanding, and verifying that code is entirely new overhead. In Yegge's words, "Reviewing enormous amounts of AI-generated code can be harder than writing it yourself."

Not every developer can work like Boris Cherny. He has years of experience that allow him to give AI clear, precise instructions. A junior developer using the same approach might miss the bugs AI introduces.

Gary Marcus Sounds the Alarm

AI critic Gary Marcus is cynical about all of this.

He recently called Matt Shumer's viral essay "weaponized hype". In response to claims that the latest AI systems build complex apps without errors, he countered: "There is not a single piece of real data to back that up."

Marcus's core argument goes like this.

"The benchmark bar is 50% accuracy. Not 100%. AI still produces hallucinations and errors."

He sees it as unlikely that AI will replace jobs within 1 to 2 years. The greater danger is executive misunderstanding. If leadership overestimates AI's actual capabilities, unrealistic expectations and irrational layoffs follow.

Is Boris Cherny's case exceptional or typical? Anthropic is an AI company. Its employees are the group most fluent in AI. Whether ordinary companies can replicate the same results is an open question.

The 50% Developer Layoff Prediction

Steve Yegge made an even bolder prediction.

"50% of Big Tech engineers will be laid off."

The logic is straightforward. If AI boosts productivity 10x, you only need 10% of the workforce for the same output. But reality is neither 0% nor 100%. So Yegge lands on 50%.

Reddit co-founder Alexis Ohanian agrees with this trajectory. Gary Marcus calls it an exaggeration.

The truth is probably somewhere in between. Not every developer will be laid off. But the gap between developers who wield AI effectively and those who do not will widen.

The reason Boris Cherny can ship 22 PRs in a day is that he has mastered how to collaborate with AI. Developers with this skill survive. Those without it fall behind.

AI Coding Adoption by Company

Here are the publicly disclosed numbers so far.

In less than a year, the AI code ratio at leading companies has doubled or tripled. If this pace holds, by the end of 2026 the majority of code at most tech companies could be AI-generated.

One caveat. The definition of "AI-generated code" varies by company. Whether it includes autocomplete, full function generation, or PR-level measurement changes the numbers significantly.

The Junior Developer Crisis

Junior developers will be hit hardest by this shift.

Traditionally, junior developers built skills by writing a lot of code. They created bugs, fixed them, got reviews, and fixed again. That cycle was how learning happened.

What happens when AI replaces this process?

If a junior developer delegates code generation to AI, they may not understand why the code works the way it does. When bugs appear, they cannot diagnose the root cause. AI fixes the code AI wrote, and the developer learns nothing.

Senior developers, on the other hand, already have strong fundamentals. When they see AI-generated code, they spot problems immediately. Experience built their judgment.

Paradoxically, the value of experience increases in the AI era. AI has leveled the playing field on code writing speed, but it has not leveled the ability to judge code.

The Future of Vibe Coding

The term vibe coding started gaining traction in late 2025. Andrej Karpathy coined the concept. It refers to building software by conversing with AI in natural language.

As of February 2026, vibe coding is no longer an experiment. At leading companies like Anthropic, Spotify, and OpenAI, it has become the default way of working.

But as Yegge warns, vibe coding carries cognitive load. AI generates code fast, and developers must review it fast. Context switching is constant. The brain gets tired.

Wikipedia now has a "vibe coding" entry. Academic research has begun. Papers on its pros and cons, productivity impact, and developer well-being are pouring in.

No verdict yet. What is clear is that there is no going back.

Is It Still Development If You Do Not Write Code

The fact that Boris Cherny has not written code in two months stirs up a range of emotions.

For some, it is liberation. Freedom from tedious, repetitive tasks. Time to focus on more creative work — design, architecture, user experience, the higher-order problems.

For some, it is loss. The satisfaction of typing code by hand, the sense of accomplishment from solving a problem directly — those things disappear.

For some, it is threat. The anxiety that your skills are becoming obsolete, that AI is coming for your job.

All three feelings are valid. And all three are real.

The New Developer Skill Set

What skills does a developer need to survive in the AI era?

First, clear instruction. Give AI vague input, get vague output. Not "build me a login feature" but "OAuth 2.0, JWT tokens, refresh token with 7-day expiry, Redis session storage." Specificity is everything.

Second, code literacy. The ability to quickly read AI-generated code and spot problems. This comes from having read a lot of code over the years.

Third, system design. AI is strong at the function level and the file level. But designing the architecture of an entire system is still a human domain. Deciding which components are needed and how they connect — that requires judgment.

Fourth, critical thinking. Never accept AI suggestions uncritically. Always ask "Why did it do this?" and "Is there a better way?" The questioning must be relentless.

Conclusion: The Definition of Development Is Changing

The fact that 90% of Anthropic's code is written by Claude is not just a statistic. It is a signal that the definition of development as a profession is changing.

Boris Cherny is no longer "a person who writes code." He is "a person who directs AI to write code." The same goes for Spotify's top developers. They give instructions on Slack, review output, and approve deployments.

Those who adapt to this shift gain unprecedented productivity. 22 PRs a day, over 50 new features shipped. Speeds that were previously impossible.

But remember Yegge's warning. 10x productivity can bring 10x fatigue. Vibe coding tops out at 3 hours a day. The human brain does not scale like AI.

The future of developers hinges on how well they work with AI. The skill of collaborating with AI is becoming more important than the skill of writing code by hand. That is the reality of 2026.

Sources:

• Top engineers at Anthropic, OpenAI say AI now writes 100% of their code — Fortune
• OpenAI's Codex and Anthropic's Claude spark coding revolution — Fortune
• Spotify says its best developers haven't written a line of code since December — TechCrunch
• The AI Vampire — Steve Yegge, Medium
• Software engineer says company leaders should expect 3 hours of vibe coding — DNYUZ
• Gary Marcus calls out viral AI essay as alarmist 'hype' — AOL
• Microsoft CEO says up to 30% of code was written by AI — TechCrunch
• The creator of Claude Code just revealed his workflow — VentureBeat
• Unsplash — Image Source