Anthropic Says AI Writes All Their Code. Science Says 29%

Let me break this down: on January 29, 2026, the internet lost its collective mind over a Fortune headline claiming that engineers at Anthropic and OpenAI no longer write code. Zero lines. The AI does everything. And here's the thing -- that headline isn't exactly a lie... but it isn't exactly the truth either.

Boris Cherny, the head of Claude Code at Anthropic, said it without flinching: "For me personally, it's been 100% for over two months. I don't even make small edits by hand." Roon, a researcher at OpenAI, confirmed the same: "100%. I don't write code anymore."

But this is where the story gets genuinely fascinating. Because between the viral headline and reality, there's a canyon-sized gap that deserves a closer look.

Boris Cherny's Numbers: 259 PRs in 30 Days

Think of it like being a film director versus a screenwriter. Boris Cherny doesn't sit down and type code. He orchestrates multiple AI agents running in parallel:

Cherny described his workflow: he opens 5 Claude Code sessions in his terminal and another 5 to 10 on Anthropic's web interface. All running simultaneously. He writes specifications, reviews outputs, approves or rejects. But he doesn't touch a single line of code.

"I've never enjoyed my job as much as I do now," Cherny told Fortune. "All the tedious work is done by Claude, and I get to be creative."

And it's not a one-off fluke. He shipped 22 PRs one day and 27 the next, each written 100% by Claude Code.

Roon (OpenAI): "Coding Always Sucked"

Imagine an OpenAI researcher publicly declaring that programming was always a necessary evil. That's exactly what Roon did:

"Coding always sucked. It was a painful prerequisite for anyone who wanted to manipulate computers to do useful things. And I'm glad it's over."

His workflow differs from Cherny's. Roon writes specifications, mentally visualizes the outcome, and launches 4 Codex instances (OpenAI's coding tool) in parallel to see multiple variations.

Here's what nobody mentions in the headlines: both Cherny and Roon work at the companies that build these tools. They use the most advanced versions, with internal access, on projects optimized for AI. This is not exactly the average developer's experience. Think of it like a Tesla engineer claiming self-driving works perfectly -- while driving on Tesla's own test track.

Dario Amodei at Davos: "6 to 12 Months"

Anthropic's CEO raised the stakes at the World Economic Forum in Davos on January 23, 2026:

"I believe we could be 6 to 12 months from the model doing most, perhaps all, of what software engineers do end to end."

But Amodei added a caveat that the headlines conveniently ignored: "I think there's a lot of uncertainty, and it's easy to see how this could take a few years."

Back in March 2025, he had already predicted AI would write 90% of code within 3 to 6 months. According to LessWrong's analysis, that prediction was "directly false as conventionally interpreted."

The LessWrong Takedown: The Real Number Is ~50%

This is where the entire narrative flips on its head.

Ryan Greenblatt, chief scientist at Redwood Research, published a detailed analysis on LessWrong that dismantles the 100% narrative:

• The actual average across all of Anthropic for merged AI-written code is approximately 50%, not 90-100%
• Some high-performing teams hit ~90%, but they're a minority
• If you include all code (throwaway scripts, disposable tools, prototypes), it approaches 90%, but that isn't what people understand when they read the headline
• The distribution is extremely uneven: 9 simple web apps at 100% AI + 1 critical algorithmic codebase mostly written by hand

Amodei himself had to walk it back later with "in many teams, not uniformly," but the media had already run with the original headline.

The Numbers That Actually Matter

Let me break this down with the data from studies that use actual methodology:

The gap is staggering. From the 100% proclaimed in the headline to the 29% measured by a peer-reviewed study in the journal Science. The reality for the industry at large sits somewhere between 30% and 50%, with outliers like Cherny at the extreme top end.

The Conflict of Interest Nobody Mentions

There's an elephant in the room that deserves your full attention.

Boris Cherny leads the product he's promoting (Claude Code). Roon works at the company selling Codex. Dario Amodei is the CEO of the company building these tools. Every primary source making the "100%" claim has a direct financial incentive to overstate the capability.

This doesn't mean they're lying. It's probably true that they personally don't write code. But extrapolating that to "AI has already replaced programmers" is a logical leap the data simply doesn't support.

As one Hacker News commenter with 500+ upvotes put it: "If it were 95% of something useful, Anthropic wouldn't have more than 1,000 employees."

What AI Still Can't Do

Andrej Karpathy -- former head of AI at Tesla and OpenAI co-founder -- is probably the most balanced voice in this debate. In December 2025, he went from 80% manual code to 80% AI code in just 4 weeks. But he also flagged real problems that anyone using these tools will recognize:

• Subtle conceptual errors: models make logical mistakes a human wouldn't
• Over-engineering: AI tends to build overly complex solutions for simple problems
• Dead code: it leaves unused functions cluttering the codebase
• Context limits: it can't hold large codebases in memory
• "Comprehension debt": accepting AI code you don't understand is a ticking time bomb

Karpathy predicts that 2026 will be the year of the "slopacolypse" on GitHub: a flood of low-quality AI-generated code swamping repositories.

And here's a genuinely alarming data point from Veracode: 29.1% of AI-generated Python code contains known vulnerability patterns. Nearly a third of the code AI produces has security holes baked in.

The METR Study: When AI Makes Experts Slower

Perhaps the most counterintuitive finding comes from the METR study, which measured the productivity of experienced developers with and without AI. The result surprised everyone -- including the researchers:

AI coding assistants decreased experienced developers' productivity by 19%.

Bain & Company reported "unremarkable" savings from generative AI in programming. And only 33% of developers trust AI outputs, according to Stack Overflow, versus 46% who actively distrust them.

Think of it like giving a professional chef a fancy new kitchen gadget. For chopping onions, it's amazing. For creating a complex sauce with 15 ingredients that need precise timing? It might actually slow them down because they spend more time fighting the tool than cooking. AI accelerates easy tasks but can genuinely slow down hard ones. Writing a CRUD function is trivial for Claude or Copilot. Designing a fault-tolerant microservices architecture... not so much.

The Real Impact on the Job Market

This is where things get serious. Because beyond the debate over percentages, real people are being affected:

Matt Garman, head of AWS, called the idea of replacing juniors with AI "one of the dumbest things I've heard." But the hiring data tells a different story.

Here's what the headlines miss: the role isn't disappearing -- it's transforming. The engineers who thrive are the ones shifting from "writing code" to "directing AI" -- reviewing, architecting, specifying. Tech lead skills and product thinking are more valuable than ever. Being a good "coder" (typing syntax) loses value; being a good "software engineer" (systems thinking) gains value.

What the Other Tech Giants Say

Interesting footnote: Microsoft is expanding internal use of Anthropic's Claude Code across its engineering teams. Not just for engineers, but also for designers and product managers.

The Emotional Dimension Nobody Wants to Discuss

There's something deeply human in this debate that goes beyond percentages and productivity metrics.

Karpathy admitted it "stings the ego a little" to program in English instead of code. Gergely Orosz, of the Pragmatic Engineer newsletter, described a feeling of "grief: something valuable is being taken away, and suddenly."

Karpathy also warned about "skill atrophy": losing the ability to write code manually. If you depend on AI for everything and one day it fails... can you solve the problem on your own?

And then there's Roon saying "coding always sucked" and that he's glad it's over. A statement that split the developer community right down the middle -- between those who felt liberated and those who felt insulted.

The Verdict: Hype vs. Reality

After analyzing 15 sources, peer-reviewed studies, and hundreds of developer comments, here's where things actually stand:

The hype: Boris Cherny and Roon genuinely don't write code. 259 PRs in a month is objectively impressive. The late-2025 models (Opus 4.5, GPT-5.2, Gemini 3) represented a real leap forward.

The reality: The 100% figure is the personal experience of two people who work at the companies that build the tools, using internal versions, on optimized projects. The real average at Anthropic is ~50%. The global average measured by Science is 29%. The gap between the headline and reality is 71 percentage points.

What actually matters: AI is genuinely transforming programming. It is not replacing it. The software engineering role is evolving from "code writer" to "AI agent director." And that's a real, significant shift -- even if it's not the apocalypse the headline is selling.

MIT Technology Review named "generative coding" as a breakthrough technology of 2026. And they're right. But the revolution is in the transformation of the craft, not its extinction.

Frequently Asked Questions

Is it true that AI writes 100% of the code at Anthropic?

Not as a company-wide figure. Boris Cherny (head of Claude Code) says it's 100% for him personally, but a LessWrong analysis revealed the real average across all of Anthropic is approximately 50%. The company itself reported a range of 70-90% specifically for Claude Code.

Is AI going to replace programmers?

Not in the sense of eliminating the job. What's changing is the role: from writing code manually to directing AI agents, reviewing, architecting, and specifying. Junior programmers are the most affected (-27.5% in employment), while seniors with systems-level skills are becoming more valuable.

How much code does AI actually write across the industry?

According to a study published in the journal Science in January 2026, 29% of new Python functions on GitHub in the U.S. are AI-generated. Microsoft reports 20-30% and Google over 30%. The figure varies enormously depending on the company, the type of code, and the tool being used.

Is AI-written code secure?

Not always. Veracode reported that 29.1% of AI-generated Python code contains known vulnerability patterns. Experts recommend mandatory human review, especially for code handling sensitive data or security logic.

What AI tools do developers use for coding?

The leading tools are GitHub Copilot (20+ million users), Claude Code (Anthropic), Codex (OpenAI), Cursor, and Lovable. Boris Cherny uses Claude Code with Opus 4.5. Roon at OpenAI uses 4 Codex instances running in parallel.