Boris Cherny, head of Claude Code at Anthropic, hasn't written a line of code by hand in over two months. The number is real. So is the gap between what's happening at AI labs and what's happening everywhere else.
Boris Cherny posted on X on a Tuesday in late January 2026, and the announcement was almost offhand: he hadn't written a line of code by hand in more than two months. The day before, he'd shipped 27 pull requests. The day before that, 22. Every single one written entirely by Claude. "I don't even make small edits by hand," he added.
Cherny runs Claude Code at Anthropic — the AI coding assistant his own team built largely by using it to build itself. He was not speaking theoretically.
The Number That Shouldn't Be Possible
Within days, the claim had spread well beyond tech Twitter. An Anthropic spokesperson filled in the context: company-wide, somewhere between 70 and 90 percent of code is now AI-generated. Claude Code itself — the product Cherny leads — was written approximately 90 percent by Claude's predecessor model. Meanwhile, an OpenAI researcher who goes by Roon on X announced that he too had stopped writing code. "100%, I don't write code anymore," he posted when asked directly. Then, in a separate post that read more like a manifesto: "Programming always sucked. It was a requisite pain for ~everyone who wanted to manipulate computers into doing useful things, and I'm glad it's over."
Klingt interessant?
The reaction was split. Some engineers celebrated; others felt the particular vertigo of watching their profession be dismissed in a sentence. Dario Amodei, Anthropic's CEO, had already sketched the trajectory — at the World Economic Forum in Davos and, days later, at a Council on Foreign Relations event in March 2025: within three to six months, AI would be writing 90 percent of the code. Within twelve months, "essentially all of it."
It's worth pausing here, because those predictions have a track record now. When Futurism checked in exactly six months after Amodei's March 2025 statement, the consensus among practitioners was that 90 percent was nowhere close. Research published in that period found that AI can actually slow some developers down — they spent less time writing code but significantly more time reviewing AI output, correcting prompts, and waiting for responses. Security researchers found that AI-generated code contained roughly ten times as many vulnerabilities as manually written code. Amodei's prediction, it turned out, was a vision statement dressed up as a forecast.
And then, four months later, Cherny posted his numbers.
The Distance Between AI Labs and Everyone Else
Microsoft's CEO Satya Nadella disclosed in April 2025 that AI was writing around 30 percent of Microsoft's code — a figure that was reported as striking. Salesforce offered a similar number. A study published in Science analyzed GitHub Python repositories and found that roughly 29 percent of functions in the United States now show signs of AI authorship, with meaningfully lower figures elsewhere.
Thirty percent versus ninety. The gap is large and it isn't accidental.
Anthropic and OpenAI are not operating like ordinary software companies. They have the most capable models in the world available not as a rate-limited API subscription but as internal infrastructure. Their engineers have spent years learning how to work with these systems — how to build context, structure feedback, and know when to intervene. Their codebases have evolved organically around AI-assisted workflows in ways that most companies haven't begun to replicate. And their internal tooling is significantly ahead of anything commercially available.
This is the honest caveat that tends to disappear in coverage of the story: the 100 percent figure is real, but it emerged under conditions that almost no other organization currently has. That doesn't make the signal meaningless. It means the figure is a limit case — useful as a reference point for where this is heading, not as an immediate benchmark for what your team should be doing next quarter.
What "100 Percent" Actually Requires
Andrej Karpathy — who has done as much as anyone to popularize AI-assisted coding through his writing and teaching — has been careful to note the constraints. Models, he observed, can make "subtle conceptual errors," over-engineer solutions, and leave orphaned code scattered through a codebase. What Cherny is doing is not passive observation. It is a new form of engineering work: less keyboard time, considerably more judgment.
The human provides architecture, direction, and scrutiny. Someone has to decide what gets built, verify that it works, and take responsibility when it doesn't. The model handles implementation — fast, tireless, indifferent to whether it's 2 AM or a public holiday. Cherny's description of the arrangement is telling: "I have never had this much joy day to day in my work, because essentially all the tedious work, Claude does it, and I get to be creative. I get to think about what I want to build next."
That isn't pure marketing. There are structural consequences that bear it out. Cherny's team now hires generalists rather than specialists, because the narrow technical skills that used to differentiate candidates matter less when the model handles implementation details. At Y Combinator, CEO Garry Tan noted in early 2025 that 25 percent of startups in the winter cohort were generating 95 percent or more of their code with AI. "That's not a typo," he wrote.
GitHub Copilot, the most widely adopted AI coding tool on the market, now counts more than 15 million users and is deployed inside 90 percent of Fortune 100 companies. GitHub's own research puts productivity gains at 51 percent faster coding. But even here the caveats are real: 29 percent of Copilot-generated Python code contains potential security weaknesses that require human review before deployment.
The picture that emerges is not one of AI replacing developers but of the profession reorganizing around a different distribution of work. The tedious parts — translating a known requirement into known syntax — are increasingly handled by models. The harder parts — deciding what to build, reviewing whether it's correct, understanding the security implications — remain stubbornly human.
What Teams Need to Decide Now
The 100 percent claim is not a target that makes sense to project onto organizations that aren't Anthropic. But as a signal of direction, it's harder to dismiss with each passing month.
For engineering teams still working in largely manual workflows, the pressing question isn't when they'll reach Cherny's numbers. It's whether their processes, architecture, and culture are positioned to move meaningfully in that direction at all. Automated tests are no longer optional — they're the precondition for deploying AI-generated code with any confidence. Precise requirements aren't bureaucratic overhead; they determine output quality more directly than the underlying model does. And the ability to review code critically — to catch the subtle errors Karpathy described — becomes the core engineering skill rather than one among many.
At nopex, we've watched companies move from skepticism to 30-plus percent AI-generated code faster than they expected, and the transition changes more than just velocity. It changes what the team does day to day, how requirements get written, and which capabilities will matter in three years. Boris Cherny isn't writing code anymore. For most engineering teams, that's not the immediate goal. But as a marker of where this is heading — and how much the profession is already changing — it's hard to look away.
