APREX · EST. 2026
AI company Anthropic, known for its Claude models, has taken an unusual and remarkable step: it is calling for a pause in the development of the most powerful artificial intelligence models. The backdrop is the company's own findings — which show that Claude is increasingly developing and solving problems without human guidance, according to Digi.no.
The Code Is Writing Itself
The figures Anthropic itself presents are striking. In May 2026, more than 80 percent of all code merged into Anthropic's codebase was written by Claude. By comparison, that figure stood in the low single digits as recently as February 2025.
As a result, Anthropic's own engineers are now delivering roughly eight times as much code per quarter as they did during the 2021–2025 period — not because they are working harder, but because Claude is doing most of the work.
From Four Minutes to Twelve Hours — and Beyond
Anthropic documents a clear development curve in the complexity and time-intensity of tasks Claude can complete autonomously. In March 2024, Claude Opus 3 was able to complete programming tasks that took a human around four minutes. A year later, Claude Sonnet 3.7 handled tasks requiring an hour and a half. By March 2026, Claude Opus 4.6 was managing tasks equivalent to twelve hours of human work.
The complexity of tasks AI can reliably complete on its own is now roughly doubling every four months.
According to Digi.no, Anthropic suggests that AI systems could handle tasks requiring several days of human effort by the end of 2026 — and tasks taking weeks by 2027.
Solved in Hours What Would Have Taken Four Years
A concrete example from April 2026 illustrates the trajectory: a Claude-based agent identified and fixed a persistent class of API errors by implementing more than 800 individual corrections. The error rate dropped by a factor of 1,000. The same task was estimated to take a human developer four years to complete.
In an internal experiment from April 2026, a Claude-based AI was tasked with carrying out an entire security research project autonomously. It formulated hypotheses, ran experiments, reported findings, and independently determined which experiments to pursue next. The result: 97 percent completion, compared to 23 percent after a week's work by human researchers. Resource usage amounted to approximately 800 cumulative work hours and the equivalent of $18,000 in compute costs.
Recursive Self-Improvement — Has a Threshold Been Crossed?
What makes Anthropic's concern particularly serious is not just isolated achievements, but the underlying trend: AI is now accelerating the development of AI. The company uses the term "recursive self-improvement" (RSI) to describe a scenario in which an AI system can design and develop its own successor entirely autonomously — without human intervention.
Anthropic's method known as "Constitutional AI" provides models with a set of written principles they can use to evaluate and correct their own responses. This means Claude already evaluates itself to a certain degree — a capability that forms part of the argument for a pause.
The Company That Created the Problem Is Asking for a Stop
There is a not insignificant contradiction in the fact that Anthropic — a company that actively develops and commercializes advanced AI — is among the first to call for a pause. The company has not published detailed requirements for what such a pause should entail, and it remains unclear what mechanism, if any, would enforce it.
Critics may point out that such an appeal from a single actor, without binding frameworks, has limited practical effect in a global and competitive AI market. At the same time, Anthropic's internal data helps substantiate the concerns with concrete figures and examples — rather than speculation.
The issue will be closely watched by policymakers and regulators around the world who are already debating frameworks for advanced AI, including the EU through the AI Act that entered into force in 2024.
Sources: Digi.no, Anthropic (internal documentation via press coverage)