@Scale
Fireside Chat with Boris Cherny, Head of Claude Code
Anthropic has seen an 8x increase in code shipped per engineer since the beginning of the year, with 80-90% of all Anthropic code now written by Claude Code.
BC
Boris Cherny
Head of Claude Code · Anthropic
41:05
The numbers, across the talk10 moments · 41:05
01 · Personal engineering output with 100% AI-written code
Anthropic
Boris pulled his own stats for the year: ~1,700 PRs, 400,000 lines added and 250,000 deleted, and 8 billion tokens used since March. Since Opus 4.5 shipped in November, 100% of his code has been written by Claude Code.
00:01:29
▲1,700 PRs
Scale
Pull requests this year
“So 1.7,000 PRs, added 400,000 wines, deleted 250,000 wines.”
00:01:29
▲400,000 lines added
Scale
Lines added
“So 1.7,000 PRs, added 400,000 wines, deleted 250,000 wines.”
00:01:29
▼250,000 lines deleted
Scale
Lines deleted
“So 1.7,000 PRs, added 400,000 wines, deleted 250,000 wines.”
00:01:29
▲8B tokens
Scale
Tokens used since March
“But since March, I used eight billion tokens.”
00:02:17
▲100%
Productivity
Code written by Claude Code
“A hundred percent of my code has been written by QuadCode since Opus 4.5.”
02 · Measuring ROI: productivity improvement per engineer
Anthropic
Boris contrasts the old dev-infra bar of 2-3% annual productivity improvement with current gains of hundreds of percentage points, citing an 8x increase in code per engineer at Anthropic since the start of the year.
00:06:12
▲8x
Productivity
Code per engineer
“And at Anthropic, you know, we've seen like an 8X increase in code per engineer since the beginning of this year.”
00:06:12
▲2-3% (before)
Productivity
Old annual productivity bar
“if you have, like, a percentage, like, 2, 3 percent productivity improvement for a year, that used to be, like, pretty good, now the productivity improvements we're looking at are, you know, like, hundreds of percentage points.”
03 · Loops: agents prompting agents to write code
Anthropic
Boris describes 'loops' as the next abstraction step above agents. On an average day roughly 30% of his code is written by loops, and on a good day he can push that to 100%.
00:11:26
▲~30% of code
Productivity
Code written by loops
“And so for me personally, maybe at this point, like 30% of my code is written by loops on an average day.”
04 · Company-wide Claude Code adoption
Anthropic
Across all of Anthropic, 80 to 90% of code is written by Claude Code on average, and for a growing share of teams — including Claude Code and Co-work themselves — it is 100%.
00:18:49
▲80-90%
Productivity
Anthropic code written by Claude Code
“I think across all of Anthropic, 80 to 90% of the code is written by quad code on average.”
05 · Optimizing token spend vs. maximizing return
Anthropic
Boris argues that cost optimization (advisor models, cheaper defaults) can cut the investment roughly 50%, while the upside from increasing return is orders of magnitude larger — so teams should focus effort on returns, not cost cutting.
00:19:38
▼~50% lower spend
Cost savings
Possible investment reductionestimated
“So like when you think about kind of ROI for something like Fable, you can reduce the investment maybe like 50% or something if you use like, you know, Fable with an advisor model or, you know, use Opus by default and have a call out to Fable when it needs.”
00:20:13
▲1,000%+ upside
Other
Opportunity to increase returnestimated
“But I actually think that if you think about R-Y, there's probably like 50% chance to reduce the investment, but probably like a thousand percent opportunity or even 100,000 or 10,000, whatever percent opportunity to increase the return.”
06 · Automated code review with Claude Code Review
Anthropic
After coding stopped being the bottleneck, code review became the next one. Anthropic built Claude Code Review, used on every internal pull request; it catches and fixes an estimated 98-99% of bugs before a human sees the PR.
00:26:04
▲98-99% of bugs
Quality
Bugs caught before human reviewestimated
“there's essentially a guarantee that all the bugs have been caught and it's not you know it's not a hundred percent we're still working on approving it but it's like you know like 98 99 percent of the bugs”
07 · CI speed-up via a dynamic workflow
Anthropic
From a single prompt, Claude Code ran a dynamic workflow orchestrating many sub-agents for a few hours, using a few million tokens, and produced four pull requests that cut CI time in half — work Boris says would have taken days, weeks or months.
00:28:05
▼-50% CI time
Cycle time
CI time reduction
“it used i think something like uh you know like like a few million tokens and it ran for like a a few hours and it produced four pro requests that reduced the i time by 50 percent”
00:28:05
▲4 PRs from one prompt
Scale
Pull requests produced
“it used i think something like uh you know like like a few million tokens and it ran for like a a few hours and it produced four pro requests that reduced the i time by 50 percent”
08 · Auto mode and prompt-injection resistance
Anthropic
Permission-prompt fatigue was hurting security, so Anthropic shipped auto mode, where a model approves or declines each action. It relies on Claude's prompt-injection resistance: per the system cards, roughly a 1% success rate at 100 attempts.
00:38:46
▼~1% at 100 attempts
Quality
Prompt-injection success rate
“but if you look at our system cards the success rate at 100 attempts is like around 1%.”