On June 4, 2026, Anthropic published a research paper titled When AI Builds Itself that made a blunt admission: more than 80% of the code merged into Anthropic’s production systems is now written by Claude. Engineers at the company ship 8× more code per day than they did in 2024. And the trend line points toward recursive self-improvement — a future where AI systems design and build their own successors without human intervention.
What Anthropic Actually Said
The post, published under the Anthropic Institute banner, lays out the case with internal data that no AI lab has publicly shared before. The argument proceeds in three stages: AI already accelerates its own development, the pace is compounding, and the gap between human and AI capability in research judgment is closing fast.
Three data points anchor the piece. First, 80%+ of merged code at Anthropic is Claude-authored as of May 2026 — up from single digits before Claude Code launched in early 2025. Second, lines of code merged per engineer per day grew 8× from 2024 to Q2 2026, largely because engineers direct and review code rather than write it. Third, in a standardized speedup test where Claude is asked to optimize a small model training script, Claude Mythos Preview achieved a 52× speedup in April 2026, compared to 3× for Claude Opus 4 a year earlier.
The post frames these findings not as a celebration but as a warning. Anthropic explicitly states: “We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for.”
The Evidence From Within
What makes the Anthropic post different from typical AI hype is the granular internal data. The company breaks down its case across engineering and research workflows.
In engineering, Claude now handles underspecified problems — ones where the engineer supplies the goal but not the method. On the most open-ended tasks, Claude’s success rate reached 76% in May 2026, up 50 percentage points in six months. One example: a routine upgrade crashed tens of thousands of training jobs, and Claude isolated the root cause in two hours — work that would normally take two to three days.
In research, Claude ran an open-ended AI safety project end-to-end in April 2026. Agents were given a problem (can a weaker model reliably supervise a stronger one?) and left to propose hypotheses, test them, and iterate. Two human researchers recovered 23% of the performance gap over one week. The Claude agents recovered 97% over 800 cumulative hours, using roughly $18,000 in compute.
The remaining human advantage, according to Anthropic, is “research taste” — choosing which problems matter and when an approach is a dead end. But on a measure of next-step judgment in real research sessions, Mythos Preview beat the human’s actual choice 64% of the time in April 2026, up from 51% for Opus 4.5 in November 2025.
Three Scenarios for What Comes Next
Anthropic sketches three possible futures. In the first, the trend stalls — capabilities plateau due to compute constraints, chip fabrication limits, or architectural ceilings. Even under this scenario, today’s models alone would reshape industries: Project Glasswing, using Mythos Preview, already found over 10,000 high-severity vulnerabilities in critical systems.
In the second scenario, AI development becomes substantially automated while humans retain direction-setting. Organizations become dramatically more efficient — “100-person companies could do the work of 10,000- or 100,000-person organizations.” Anthropic calls this the most likely near-term outcome.
In the third, full recursive self-improvement arrives: AI systems design and train their own successors with minimal human oversight. Anthropic warns this would increase risks of humans losing control over AI systems, making security, monitoring, and behavior-shaping mechanisms far more critical.
How the Industry Reacted
The post triggered immediate coverage. Interesting Engineering compared the moment to science fiction becoming real. Axios highlighted Anthropic co-founder Jack Clark’s call for lawmakers to understand what is coming. The Economic Times reported that Anthropic wants AI development to slow down globally.
The paradox is hard to miss: Anthropic is both accelerating AI development by delegating more work to Claude and publicly calling for institutions to prepare for the consequences of that acceleration. As the post itself notes, “The rate at which organizations can spot and fix these bottlenecks may be a skill that determines competitive advantage.”
Why This Matters for Everyone
The implications extend well beyond AI labs. If recursive self-improvement becomes real, the gap between organizations that adopt AI-powered development and those that don’t will widen into a canyon. Anthropic’s own data shows that employees using Mythos Preview produce roughly 4× more output than they would without AI tools.
For software teams, the message is direct: code review, not code writing, is becoming the bottleneck. When an AI system can generate more code than humans can review, the entire engineering workflow inverts. Organizations need to invest in review infrastructure, automated testing, and architectural decision-making — the skills that remain human-domain, for now.
For security teams, the implications are double-edged. AI that can build AI can also find vulnerabilities at scales humans cannot match. Project Glasswing’s 10,000+ vulnerability discoveries prove this. But the same capability, pointed at offensive operations, represents a fundamental shift in the threat landscape.
For policymakers, Anthropic’s warning is unambiguous: “It could come sooner than most institutions are prepared for.” The company is effectively asking governments to catch up — not with regulation that stifles innovation, but with institutional capacity to monitor, verify, and respond to increasingly autonomous AI systems.
Key Metrics at a Glance
| Metric | 2024 / Early 2025 | May–June 2026 |
|---|---|---|
| Code authored by Claude | Low single digits % | 80%+ of merged code |
| Code per engineer per day | Baseline (1×) | 8× increase |
| Model optimization speedup | 3× (Opus 4, May 2025) | 52× (Mythos Preview, April 2026) |
| Open-ended task success rate | 26% (Nov 2025) | 76% (May 2026) |
| Next-step judgment vs. human | 51% (Opus 4.5, Nov 2025) | 64% (Mythos, April 2026) |
| Employee productivity with Mythos | Baseline | ~4× self-reported output |
| Research gap recovery (agents) | N/A | 97% vs 23% human baseline |
The Deeper Question Nobody Answered
Anthropic’s post is notable for what it does not resolve. The company acknowledges that “research taste” — the judgment to choose which problems matter — remains the last meaningful human advantage. But the data shows this gap closing: from 51% to 64% in five months on the next-step judgment metric.
If that trend continues at the same rate, the crossover point arrives before the end of 2026. At that moment, the question shifts from “can AI help build AI?” to “what role is left for humans in AI development?”
Anthropic’s own employees are grappling with this. The post includes candid quotes from staff: “On days where everything works well, I can’t help but think nothing I do matters, everything is automated and better and faster than I ever will be.” And from another: “It’s been ~5 months since I last wrote any code myself.”
The post ends where it began: with a call for preparation, not panic. But the data it presents makes one thing clear — the window for preparation is closing faster than most people think.
Sources
- Anthropic Institute — When AI Builds Itself (June 4, 2026)
- Interesting Engineering — Anthropic says AI could soon create more advanced versions of itself
- Axios — Anthropic warns AI could soon help build its own successors
- Economic Times — Anthropic wants AI development to slow down globally
- Kingy AI — Inside the Recursive Self-Improvement Race
- Indian Express — What happens when AI starts building AI