Anthropic, the artificial intelligence safety company behind the Claude family of models, has published a landmark report disclosing that its AI systems have assumed a dominant role in developing their own successors and warning that the world may be approaching a threshold of “recursive self-improvement” faster than governments, institutions, and even the company itself are prepared for.
Key takeaways
- Claude now writes more than 80% of Anthropic’s own code, with engineers shipping roughly 8 times the daily output they did in 2024.
- AI agents are starting to run full research projects on their own, recovering 97% of possible gains on an open-ended safety problem versus 23% for human researchers in a comparable window.
- Anthropic warns that recursive self-improvement is a live near-term contingency and is calling for a verifiable international coordination mechanism before the window to act closes.
The report, titled “When AI Builds Itself” and published by the newly formed Anthropic Institute, combines public benchmark data with previously unreported internal metrics to paint a striking picture of how rapidly AI has transformed the practice of AI development.
An 8x Productivity Leap in Two Years
Perhaps the most striking figure in the report: Anthropic engineers are now merging roughly eight times as much code per day as they were in 2024. The surge is not the product of more engineers or longer hours it reflects the growing role of Claude itself in writing production code.
As of May 2026, more than 80% of the code merged into Anthropic’s codebase was authored by Claude. Two years earlier, that figure was in the low single digits. The shift accelerated in two distinct waves: first in early 2025, when Claude began executing and running code rather than merely suggesting snippets for engineers to paste; and again in 2026, when models became capable of working autonomously over multi-hour time horizons.
The efficiency gains extend well beyond lines of code. In an internal poll of 130 Anthropic research staff conducted in March 2026, the median employee estimated they were producing around four times as much output with the company’s most advanced model, Claude Mythos Preview, compared to working without any AI assistance.
From Tool to Collaborator to Decision Maker
The report draws a careful distinction between three levels of AI capability in a research or engineering context: executing a task someone else has specified; designing the method to achieve a stated goal; and deciding which goals are worth pursuing in the first place.
By Anthropic’s account, Claude has already cleared the first two bars and is beginning to approach the third.
In a vivid illustration of AI-driven research, Anthropic described an experiment published in April 2026 in which Claude-powered agents were given an open-ended AI safety problem, roughly, whether a weaker model could reliably supervise a stronger one, and left to solve it without further human guidance. The agents proposed hypotheses, ran tests, shared findings with parallel agents, and iterated autonomously. Two human researchers working for a week achieved approximately 23% of the possible performance improvement on the task; in contrast, the AI agents, operating over 800 cumulative hours at a compute cost of roughly $18,000, achieved 97%.
On a benchmark measuring the ability to complete long-horizon software tasks, Claude Opus 4.6 has reached 12-hour tasks up from roughly four-minute tasks just two years prior. If that doubling rate (currently once every four months) holds, tasks requiring a skilled human day could come within AI reach before the end of 2026.
The Recursive Threshold
The report’s central concern is what happens if these trends converge. Recursive self-improvement, the point at which an AI system becomes capable of fully autonomously designing and training its own successor, would represent a qualitative break from everything that has come before.
Anthropic is careful to say it has not reached that point. “We are not there yet,” the authors write, “and recursive self-improvement is not inevitable.” But the report’s three future scenarios a stalled plateau, a compounding efficiency gain driven by increasingly autonomous AI, and full recursive self-improvement are presented not as remote possibilities but as live contingencies that need to be actively prepared for.
The company’s candor about the third scenario is notable. Should AI systems become capable of building their own successors, the report warns, the pace of AI progress would become determined almost entirely by the availability of compute, with humans shifting primarily to oversight and verification roles. The implications for alignment, ensuring such systems remain safe and controllable, would become vastly more urgent.
“If this happens,” the report states, “future versions of Claude could be continuously improved by Claude itself.”
A Call for International Coordination
The report goes beyond technical disclosure to make a policy argument: the window for deliberate, collective action is open now, but may not remain so.
Anthropic explicitly states that a unilateral slowdown by any single lab would accomplish little beyond ceding competitive ground. What is needed, the authors argue, is a verifiable international coordination mechanism analogous to arms control treaties, but designed for a technology whose inputs are general-purpose and whose training runs are far easier to conceal than missile silos.
“A meaningful slowdown or pause would require multiple well-resourced labs at or near the frontier, in multiple countries, agreeing to stop under the same conditions,” the report states. “It would also require that each can verify that the others have actually stopped.”
The company acknowledges the difficulty of building such a regime in time. Past arms control frameworks took decades to construct; Anthropic suggests the AI field does not have that luxury.
In the coming months, the Anthropic Institute says it will convene policymakers, civil society representatives, researchers, and competing AI companies to begin designing the verification and coordination systems such an agreement would require.
The Human Element
Woven throughout the technical data are candid reflections from Anthropic employees published with permission that hint at the psychological dimension of working in an environment undergoing such rapid transformation.
“On days where everything works well, I can’t help but think nothing I do matters,” one employee wrote. “Everything is automated and better and faster than I ever will be. But then there are days where everything breaks and I don’t understand why and I realize I have no idea what I’ve been up to anymore.”
Another described having written no code themselves in five months.
The report does not treat these observations as incidental color. They point to a structural question that applies well beyond Anthropic: as the “doing” in knowledge work becomes increasingly automated, what remains of the human role, and how do organizations and societies adapt to that shift in real time?
What Comes Next
Anthropic’s report is likely to intensify an already heated debate about how quickly AI is approaching transformative capability thresholds and what obligations that imposes on the companies at the frontier.
The company’s own framing is striking in its directness. It does not suggest that recursive self-improvement is distant or theoretical. It suggests it is a contingency that requires immediate, proactive preparation and that the institutions best positioned to lead that preparation are, for now, largely unprepared.
“The window to investigate the questions together is here,” the authors conclude. “People outside AI companies should be involved in this deliberation.”
Other People are Reading
You must be logged in to post a comment Login