As of May 2026, more than 80% of the code merged into Anthropic’s production codebase was written by Claude — up from low single digits just 15 months ago. The company published a paper on June 4 warning that AI is already accelerating AI development, with recursive self-improvement arriving sooner than most institutions are prepared for.
Anthropic’s research paper, “When AI Builds Itself”, published Thursday alongside a striking productivity number: engineers at the company are now merging 8 times as much code per day as they were in 2024. That’s not a small team working harder — that’s the same people, with Claude doing most of the typing.
From 4% to 80% in 15 Months
When Anthropic’s in-house coding agent launched in February 2025, Claude contributed only low single-digit percentages of merged code. By May 2026, that figure crossed 80%. The jump happened without a headcount explosion — it happened because Claude got dramatically better at the hardest parts of the job.
On the most difficult, least-specified coding tasks, Claude hit a 76% success rate in May 2026, a gain of 50 percentage points in just six months. On raw code optimization benchmarks, Claude achieved a 52x speedup compared to 4x for a skilled human engineer. Engineers are now merging 8x more code per day than they were in 2024.
Anthropic AI Recursive Self-Improvement: How Close Is It?
Recursive self-improvement is when an AI system meaningfully accelerates the development of its own successors — reducing the role of human engineers and compounding each generation’s gains. Anthropic’s Mythos Preview model already picks the better next step 64% of the time (up from 51% for Claude Opus 4.5) when shown a session where a researcher took a wrong turn.
Anthropic is explicit about its current status: “We are not there yet, and recursive self-improvement is not inevitable.” But the June 4 paper adds the threshold could arrive sooner than most institutions are prepared for — and calls it worth planning around now. Tom’s Hardware and Scientific American both covered the recursive self-improvement warning in depth.
Anthropic Wants a Global Pause Button
In the same paper, Anthropic called for the world to have “the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up.” This isn’t a ban proposal — it’s an emergency brake, a coordinated international mechanism that could be triggered if development outpaces governance. The company framed the pause option as a safeguard, not a surrender: something the world should have available before it’s urgently needed.
For a company that profits from shipping faster, that ask is notable. It’s a direct acknowledgment that the 80% code figure isn’t just a productivity milestone — it’s a signal about what’s coming next. See how Claude compares today in our Claude vs ChatGPT 2026 comparison.
Frequently Asked Questions
Is Claude undergoing recursive self-improvement right now?
No. Anthropic explicitly states it has not reached recursive self-improvement. Claude is writing production code for Anthropic’s systems, but engineers still oversee all major architectural decisions. The 80% figure reflects code volume, not autonomous model development.
What is recursive self-improvement in AI?
Recursive self-improvement is when an AI system meaningfully accelerates development of its own successors — reducing the need for human engineers to drive progress. Anthropic says this threshold hasn’t been crossed, but the trajectory makes it worth planning for now.
How fast did Claude’s code contribution grow?
Claude went from low single-digit percentages in early 2025 to over 80% of merged production code by May 2026 — roughly a 15-month transformation. Engineers are now merging 8x more code per day than they were in 2024, without adding headcount.
What is Anthropic’s “global pause button” proposal?
Anthropic is calling for an internationally coordinated mechanism to slow or temporarily halt frontier AI development if progress outpaces safety research or governance capacity. It’s framed as an option, not a mandate — an emergency brake the world should have available.
What does this mean for software engineers?
Anthropic’s engineers are merging 8x more code per day than in 2024 — AI isn’t replacing them, it’s amplifying them. The nature of engineering work is shifting fast toward oversight, direction, and review. Our best AI coding assistants for 2026 covers the tools driving this shift.
Anthropic’s “When AI Builds Itself” is one of the clearest windows yet into what AI-accelerated development looks like inside a frontier lab. Also worth reading: our breakdown of the latest Claude Fable 5 release for context on where Claude’s capabilities stand today.
Last Updated: June 2026
