Anthropic is delegating a growing share of AI development to AI systems themselves — and the data from inside the company suggests the pace is accelerating faster than most institutions are prepared for
What the numbers show
As of May 2026, as reported here, Claude authors more than 80% of the code merged into Anthropic’s codebase — up from single digits before Claude Code launched in early 2025. The result: engineers are merging roughly 8× more code per day than they did in 2024, not because they’re working harder, but because Claude is writing most of it.
External benchmarks confirm the same trend. On SWE-bench, a real-world software engineering test, models went from single-digit scores to near-saturation in two years. On CORE-Bench, which tests whether AI can reproduce existing research, saturation came in 15 months. METR’s task-horizon research shows the length of tasks AI can reliably complete has been doubling roughly every four months.
From assistant to researcher
Claude isn’t just writing more code — it’s increasingly making research decisions. In April 2026, Anthropic published a demonstration of Claude-powered agents running an open-ended AI safety research project end to end: proposing hypotheses, testing them, and iterating — recovering 97% of the performance gap that two human researchers recovered only 23% of in a week.
Separately, analysis of 129 real research sessions found that Claude Mythos Preview suggested a better next experimental step than the human researcher 64% of the time — up from 51% just six months earlier.
The gap between AI today and a fully autonomous AI researcher is narrowing to a single dimension: the judgment to choose which problems are worth working on at all.
Why the urgency to slow down?
The concern isn’t that AI turns “evil” — it’s that things move so fast and become so complex that humans gradually lose the ability to understand, steer, or reverse course. Four specific risks drive this fear:
Small flaws in today’s models could get baked into the next model they help build, then amplified further. Each generation inherits and potentially magnifies the errors of the previous one — until the behavior is far from what humans intended and too complex to trace back.
If AI generates more code and research than humans can meaningfully review, oversight stops being real. You’re technically “in control” but practically rubber-stamping outputs you can’t fully evaluate.
Laws, safety frameworks, and international agreements take years to develop. If AI capabilities double every few months, societal institutions fall further behind with each cycle — and by the time they catch up, the technology may already be operating at a scale that’s very hard to constrain.
Once AI is good enough to meaningfully improve its own successor, each new generation arrives faster than the last. The concern isn’t just where we are today — it’s that the rate of change itself accelerates, leaving even less time to course-correct if something goes wrong.
Three possible futures
Progress plateaus as returns on scale diminish or supply-chain constraints (energy, chips) bite. Even frozen at current capability, models like Claude Mythos Preview are already reshaping industries — finding over 10,000 critical software vulnerabilities in Project Glasswing’s first weeks.
AI handles most execution while humans steer direction. Organizations scale dramatically — a 100-person company doing the work of 10,000. Human bottlenecks (code review, prioritization) become the new constraint, per Amdahl’s law. Anthropic considers this the most likely near-term scenario.
AI systems design and refine their own successors with minimal human involvement. Progress is gated only by compute availability. How alignment problems are resolved — or not — in this scenario is the question Anthropic is least certain about.
What Anthropic says should happen next
Anthropic acknowledges that a meaningful slowdown in frontier AI development would “likely be a good thing” — but only if multiple well-resourced labs across multiple countries pause together under verifiable conditions. A unilateral pause simply changes who leads; it doesn’t create the deliberative process that is currently missing.
Building that verification infrastructure is the explicit goal of the Anthropic Institute, which plans to convene policymakers, researchers, civil society, and other AI companies in the coming months. The full analysis and agenda are published in the original report.
A race with no brakes — and no referee
Anthropic is calling for a slowdown, but already knows how unlikely it is to achieve. Stopping unilaterally would mean ceding ground to those who care less about safety. And convincing everyone else to stop together would require a level of international cooperation that simply doesn’t exist today.
The structure of the problem increasingly resembles a technological cold war: two main powers — the United States and China — competing for a technology that could reshape global balances, aware of the risks but unable to escape the logic of competitive advantage. With one crucial difference from the nuclear arms race: back then, the risks were visible and the main actors were states. Today the technology is also developed by private companies, is impossible to verify from the outside, and its effects are gradual — with no equivalent of Hiroshima to force the world to stop and reckon.
We are probably in the most dangerous phase of this analogy: the 1950s, before treaties, diplomatic hotlines, or shared deterrence doctrines existed. The rules of the game have not yet been written. The question is not whether an international coordination mechanism will be needed — it’s whether there will be enough time to build one.
