Anthropic has released a research paper titled "When AI builds itself" that delivers both a remarkable milestone and a sobering warning: Claude now authors more than 80% of the code merged into Anthropic's production codebase, and the company believes full recursive self-improvement — where AI autonomously designs, trains, and deploys its own successors — could be closer than most assume.
Key Highlights
- Claude wrote over 80% of all code merged into Anthropic's production repositories in May 2026
- Typical engineers at Anthropic are now merging eight times as much code daily compared to 2024 levels
- Anthropic's Mythos Preview model can handle tasks spanning at least 16 hours — beyond current benchmark upper limits
- The company calls for a verifiable global mechanism that would allow multiple frontier labs to simultaneously pause frontier AI development
- The paper draws an explicit parallel between AI governance and nuclear arms control treaties
Claude's Code Milestone
The numbers are striking. Before Anthropic launched Claude Code in early 2025, Claude contributed only a small fraction of the company's codebase. By May 2026, that figure had surpassed 80% of all merged production commits. Engineers are not being replaced — they are being dramatically amplified. According to the paper, the typical Anthropic engineer is now shipping eight times more code per day than they did in 2024.
The Mythos Preview model, Anthropic's most advanced internal system, is already handling tasks that span at least 16 continuous hours — a duration that exceeds what current AI benchmarks are even designed to measure.
The Recursive Self-Improvement Risk
The paper's central concern is what the authors call recursive self-improvement (RSI): the scenario in which an AI system begins designing, building, and training its own successors with minimal human involvement. As Anthropic describes it, humans would be pushed to the margins of a process they currently run from start to finish.
The risk is not simply that AI becomes more capable. It is that any subtle misalignment present in one generation of models could compound across generations — growing more frequent and less understood with each cycle, until humans lose the ability to course-correct.
"If it were possible to effectively slow the development of this technology to give ourselves more time to deal with its immense implications, we think that would likely be a good thing," the paper states.
A Conditional Call for a Global Pause
Rather than calling for a unilateral halt, Anthropic proposes a conditional global mechanism: a verifiable agreement in which multiple frontier AI labs across multiple countries commit to pausing development simultaneously, under identical conditions that can be independently audited. The model, the paper notes, mirrors nuclear arms control frameworks.
The proposal is deliberately conditional. Anthropic makes clear it would not unilaterally stop development while competitors continue advancing. Critics have noted the obvious tension: the same company now preparing for an initial public offering — with Anthropic's valuation approaching one trillion dollars — is simultaneously urging the industry to consider pumping the brakes.
IPO Timing and Industry Reaction
Anthropic filed confidential IPO paperwork last week, and the timing of the "When AI builds itself" paper has attracted attention. Some observers read the publication as a strategic move — reinforcing Anthropic's safety-first positioning ahead of a landmark public offering. Others argue the underlying research is genuine and the concerns are well-founded regardless of corporate context.
The paper has generated over 3,700 posts on X and sparked debate among researchers, engineers, and ethicists about whether the industry is moving too fast to govern itself.
What's Next
The gap between where AI systems are today and full recursive self-improvement is not fixed. As Anthropic's own data shows, capabilities are accelerating faster than many anticipated. Whether or not the industry converges on a pause mechanism, the "When AI builds itself" paper marks a significant moment: one of the world's most prominent AI labs has publicly acknowledged that the feedback loop between AI and AI development is already underway.
Source: The Next Web