The person behind the numbers
Boris Cherny is the creator and Head of Claude Code at Anthropic. Before that he spent five years at Meta as a Principal Engineer, where he owned code quality across the entire company. He knows what it looks like when hundreds of engineers grind for a year and move a productivity metric by a few percentage points. That experience makes his current numbers land differently.
Since last November, 100% of Cherny's code has been written by Claude Code. He uninstalled his IDE. He has not manually edited a single line. This year he shipped roughly 1,700 pull requests, adding 400,000 lines and deleting 250,000. He has consumed over 8 billion tokens since March. Speaking at Fortune Brainstorm Tech in June 2026, he said: "I haven't written a line of code by hand in, I think, eight months now... Claude Code, 100% written by Claude Code."
His own workflow has also changed in a more fundamental way. "I don't prompt Claude anymore," he said at a Scale AI fireside chat in June 2026. "I have loops running that prompt Claude and figuring out what to do. My job is to write loops." That's not a productivity tip. It's a different job description.
What Anthropic's numbers actually show
Anthropic has doubled its engineering org since adopting Claude Code internally. According to Cherny on Lenny's Podcast, per-engineer productivity measured in merges per engineer per day has increased 200%. Separately, at the Scale AI fireside, he cited an 8x increase in code per engineer since the start of the year. Anthropic's own blog post added a direct caveat to the larger figure: it was "almost certainly an overstatement," because measuring lines of code rewards volume, not quality.
That caveat is worth taking seriously. But even the more conservative 200% figure represents a different order of magnitude from anything Cherny saw at Meta. Across Anthropic, 80-90% of code is now written by Claude on average. For a growing share of teams, the number is 100%.
One quieter data point says something about ramp-up quality: new engineer onboarding dropped from weeks to roughly two days. New hires ask Claude how to query a database, and Claude already knows the codebase well enough to answer accurately.
The bottleneck migration problem
The most practically useful thing Cherny talks about is what happens after you automate code generation. Output did not just scale; it shifted where friction accumulates.
When code generation stopped being the bottleneck, code review became one. Anthropic's response was to automate that too: a team of Claude instances with distinct personas collaborates to review every pull request, catching "pretty much every bug." A human still approves the final merge, but Claude does the review work.
Then maintainability and security emerged as the next constraints. Anthropic now runs automated Claude-driven routines that iteratively improve the codebase, plus a Claude Security product that scans for vulnerabilities on a rolling schedule.
This sequence is the actual playbook: identify where friction moved, then automate that stage, then repeat. Cherny described the eventual state at the Scale AI fireside: "Once you get it to this point... The bottleneck is going to be good ideas."
His advice for companies that haven't started yet
Most companies respond to rising AI token costs by clamping down on access. Cherny's argument runs the other direction, and he made it plainly at the Scale AI event: "The way to do this is give people tokens and give them safety to experiment so they feel like they can try stuff and they're not going to get penalized for it."
His reasoning is that the best internal use cases often come from unexpected places. An accountant in a corner of the org, a marketing person the CEO has never heard of. Restricting access to a core engineering group means those ideas never surface.
Once promising use cases emerge, he recommends shifting cost control to the back end. "Once you find these internal use cases that kind of work, then you want to control the costs and you want to do that on the backend, not on the front end." That's a meaningful operational distinction: you're not gatekeeping access, you're managing spend after adoption proves value.
On ROI framing, he pushes back against comparing Claude Code to existing developer tools. "Compare it to what the cost would have been if an engineer had done this work." The comparison class matters. If you benchmark it against a coding assistant, the returns look incremental. If you benchmark it against fully loaded engineering cost, the math changes.
For building the internal business case, he recommends a simple pilot: have one team use Claude Code and another work without it, then measure differences in speed, security, and output quality. Let those numbers make the argument rather than asking leadership to take it on faith.
What this looks like at the hiring level
Cherny's team structure reflects the same philosophy. He hires generalists over specialists, values low-ego collaborators, and looks for empiricists who defer to customer data rather than internal conviction. "We treat everyone on the team as essentially a CEO," he said at Fortune Brainstorm Tech.
That framing makes sense given how the work has changed. If your job is designing autonomous loops rather than writing functions, you need people who can reason about systems, handle ambiguity, and make judgment calls about what to build. Deep specialization in any one language or framework matters less than it did two years ago.
Claude Code reached $1 billion in annualized revenue within six months of its May 2025 general availability launch. As of February 2026, it authored approximately 4% of all public GitHub commits, roughly 135,000 per day. The tool grew 42,896x in 13 months from its research preview. Those numbers describe adoption velocity. The more interesting question, which Cherny is trying to answer in public, is whether other organizations can get their engineering orgs to actually change how they work rather than just adding a new tool to an existing workflow.