Roughly 90% of the code inside Anthropic is now written by Claude Code, according to the team that builds it. This article is not a hot take or a vendor pitch. It is an aggregation of every public stat the Claude Code team has dropped over the last six months, with the receipts attached.
These numbers come from podcast appearances, conference talks, and the X accounts of the people running the product. They are self-reported, which is exactly why they are interesting. The team is putting them in writing.
The Headline Number 90% of Anthropic's Code
Cat Wu, who leads product for Claude Code, said it plainly in Everyone's an Engineer Now on O'Reilly Radar: 90% of Anthropic's own code is now written via Claude Code. That is the company that designed the model, the harness, and the workflow, reporting that nine out of every ten lines of internal code came through the tool they ship to customers.
Sid Bidasaria, an engineer on the Claude Code team, confirmed a similar figure on the MLOps Community podcast. By his estimate roughly 90% of Claude Code's own codebase is Claude-written, and test code is closer to 100%. The framing matters. Tests are predictable and well-scoped, which is exactly the kind of work where the model excels. Production logic still gets reviewed and edited by humans, but the typing has shifted almost entirely to the model.
This is not a vendor case study. It is the team that built the tool reporting on themselves, which is the closest thing you get to an unvarnished benchmark in this industry. They are not selling licenses to themselves.
The Claude Code team is not just shipping a product, they are the heaviest internal users of it. The stats below are the receipts they post publicly, not marketing claims. Every number links to its primary source.
The remaining sections walk through the productivity multipliers, the individual records, and the external case studies from Spotify and the rest of the customer roster. The throughline is that these numbers keep showing up across different talks and tweets, which makes them harder to dismiss as a one-off marketing moment.
Productivity Per Engineer Doubled
Boris Cherny, who leads Claude Code, has been the most public about per-engineer numbers. On the YC Lightcone podcast in February 2026, he reported a 150% per-engineer productivity gain at Anthropic, with 70 to 90% of internal code AI-generated. That spread is honest. Some teams are at the high end, others are still ramping, and the average is what gets cited externally.
A month later, on the March 9 2026 launch of Code Review, Boris wrote that code output per Anthropic engineer is up 200% this year and that reviews had become the bottleneck. That last detail is important. The constraint shifted from writing code to reviewing it, which is why the team prioritized building review tooling next.
Cat Wu added another data point on February 11 2026 with the launch of contribution metrics: a 67% increase in PRs per dev at Anthropic from using Claude Code. Three different numbers, three different speakers, three different framings. They are not the same metric, but they all point in the same direction, and they keep appearing in different contexts over a six-month window.
The pattern here is not that Anthropic engineers got smarter or worked longer hours. It is that the same headcount is moving more code per week, and the team is being transparent about the multiplier. The next section moves from team averages to one specific engineer's logs, which is where the numbers stop feeling abstract.
Boris's Personal Numbers 259 PRs in 30 Days
On December 27 2025, Boris posted his thirty-day stats on Threads.
In the last thirty days, I landed 259 PRs, 497 commits, 40k lines added, 38k lines removed. Every single line was written by Claude Code + Opus 4.5.
That is roughly nine merged PRs per workday from one engineer, with the typing fully delegated. Two months later on Lenny's Podcast, he doubled down.
100% of my code is written by Claude Code. I haven't touched a line by hand since November.
The Sequoia AI Ascent 2026 keynote added the daily ceiling: roughly 150 PRs landed in a single day at his personal peak.
These are extreme numbers, and they come with an obvious caveat. Boris designed the workflow, built the harness, and is shipping in a codebase he knows intimately. But the absolute volume is still worth sitting with. One engineer, 259 PRs a month, zero lines typed by hand. Even at a steep discount that is a different operating point than most teams have ever seen.
We're publishing a deep-dive series on the doctrine and tooling the Claude Code team uses every day.
Browse our Claude Code coverageThe natural pushback is that these are insiders measuring insiders. So the next section looks at what happens when external companies adopt the same workflow at scale, starting with the one that ended up in Spotify's earnings call.
Outside Anthropic Spotify and the Customer Case Studies
On February 13 2026, Boris posted about Spotify's results.
Their best developers haven't written a single line of code since December, they fix bugs from their phones, and they shipped 50+ features from Slack during morning commutes.
The same week, Cat Wu put a sharper frame on the same shift.
Spotify's best engineers are supervising Claude Code instead of writing code manually.
That phrase, supervising instead of writing, is the most useful framing in this whole article. It is not that the engineers stopped engineering. It is that the unit of work shifted from typing a line to reviewing a PR, kicking off a session, or describing a fix. The Slack commute story is the operational version of the same idea. A senior engineer can dispatch a fix from their phone because the loop is mature enough that they trust the output.
The same pattern appears across the public customer roster. Ramp, Rakuten, Brex, Wiz, and Shopify have all been cited by Anthropic as production users with similar workflow shifts. The team uses what they ship, and so do their biggest customers, and the case studies are starting to converge on the same shape. Heavy CLAUDE.md investment, plan mode by default, custom skills, and aggressive use of subagents.
Here is the same evidence laid out in order so the cadence is visible at a glance.
| Date | Source | Number |
|---|---|---|
| Dec 27, 2025 | Boris Cherny (Threads) | 259 PRs in 30 days, 100% Claude-written |
| Feb 11, 2026 | Cat Wu (contribution metrics launch) | 67% more PRs per dev at Anthropic |
| Feb 13, 2026 | Spotify Q4 2025 earnings | Best engineers supervising Claude Code |
| Feb 17, 2026 | Boris Cherny (YC Lightcone) | 150% per-engineer productivity gain |
| Feb 19, 2026 | Boris Cherny (Lenny's Podcast) | 100% of his own code is Claude-written |
| Mar 9, 2026 | Boris Cherny (Code Review launch) | 200% code output year over year |
| Apr 2026 | Cat Wu (O'Reilly Radar) | ~90% of Anthropic's code is Claude-written |
The timeline matters because it shows a steady drip of numbers rather than a single splashy announcement. December gave us Boris's personal stats. February brought three separate launches and the Spotify story. March added the 200% output figure. April put the 90% number into a published O'Reilly piece. None of these are anomalies. They are a sustained pattern.
Reading these numbers as a marketing pitch. The team has been transparent about the caveat. These numbers come from people who designed the workflow, share a CLAUDE.md culture, and have unlimited token budgets. Your mileage scales with how much of that you replicate.
The caveat is real, but it cuts both ways. Most teams will not hit 90%. Many will hit 30 or 40 with a moderate investment in tooling and conventions. The interesting question is not whether the numbers transfer cleanly. It is what they tell us about the ceiling.
Why the Numbers Matter Beyond Anthropic
The honest read is that these stats matter because they show what is possible when the workflow is right, not because every team will replicate them on day one. The Claude Code team has unlimited token budgets, a deep institutional CLAUDE.md culture, and a feedback loop with the model itself. Most teams have none of those advantages on day one.
But the ceiling has moved. Before these numbers, the upper bound on AI-assisted development was a vague but conservative estimate. After them, the upper bound is documented at roughly one engineer landing nine merged PRs a day with zero hand-typed code. That reframes the planning conversation for every team building software.
The other thing these numbers do is set up the rest of this series. The doctrine pieces, plan mode discipline, CLAUDE.md as constraints, verify-app skills, are what closes the gap between an average team's results and the Claude Code team's results. The stats are the upper bound. The workflow is how you climb toward it.
What This Means For You
- If you're a founder: These numbers reframe what AI-assisted means at the upper bound. The ceiling is higher than most planning conversations assume, and pricing decisions about token budgets versus headcount look very different when one engineer can land 259 PRs a month.
- If you're changing careers: The team's bet is that the cost of execution drops fast. Ruthless prioritization, clear specifications, and review judgment replace typing speed as the rate-limiting skill, which is a friendlier shift for career-changers than most assume.
- If you're a student: Read the primary sources, not the recap pieces. The team posts the raw numbers themselves on X, on Threads, and in podcast transcripts. Following Boris, Cat, and Sid directly gives you a cleaner signal than any aggregator.
Plan mode, a real CLAUDE.md, and a verify-app skill. Three things to do today.
See the team's playbook