When AI Builds Itself: Inside the Numbers That Made Anthropic Ask for a Pause
Anthropic's own data shows Claude now writes 80% of the company's code. One engineer hasn't written a line in five months. The company that built Claude is asking the world to pause.
In February 2025, Claude wrote roughly 1% of Anthropic’s own code. By May 2026, that number had crossed 80%. The code that humans write at Anthropic is no longer the majority of what ships. It is the shrinking remainder.
That is one of the central revelations in “When AI builds itself,” a report released June 4 by the Anthropic Institute, the research arm of the company behind Claude. The document is the most detailed public accounting yet of how far AI has moved into the territory once reserved for human engineers and researchers. And it is the first time a frontier lab has released this level of internal data about its own operations.
The report does not pretend to be neutral. It contains a call for a global pause on training systems beyond a certain capability threshold. The company that built Claude is asking the world to slow down, and it is using its own numbers to explain why.
The Engineer Who Stopped Writing Code
One of the most striking passages in the report comes from an anonymous Anthropic employee. “I started leaning hard into Claudifying about a year ago,” they told the researchers. “That’s been a crazy adventure and it’s now been roughly 5 months since I last wrote any code myself.”
Five months without writing code, at a company that builds frontier AI models. The person is still an engineer. They still review, direct, and approve. But the doing itself has migrated to the system they helped create.
This is not an outlier. Across Anthropic, the median engineer now ships eight times as much code per quarter as they did between 2021 and 2024. The report is careful to note that lines of code is a flawed metric, and the agency admits the true productivity gain is certainly lower than 8x. But the direction is unambiguous. In a March 2026 poll of 130 employees across Anthropic’s research teams, the median respondent estimated they produced roughly 4x as much output with Claude Mythos Preview as they would have without it.
The question the report raises but does not fully answer: what does “engineer” mean when the engineer no longer engineers? The role is shifting from builder to director. The person who hasn’t written code in five months is still valued, but the nature of that value has changed. They set goals. They review work. They make judgment calls about direction. The execution is handled elsewhere.
That shift is not painless. Another employee quoted in the report put it bluntly: “On days where everything works well, I cannot help but think nothing I do matters, everything is automated and better and faster than I ever will be. But then there are days where everything breaks and I do not understand why and I realize I have no idea what I have been up to anymore.”
The 800 Bugs That Would Have Taken Four Years
The report includes a number that deserves attention: in April 2026 alone, Claude shipped over 800 fixes that reduced a class of API errors by a factor of one thousand. The engineer overseeing this work estimated that a human would have needed four years to complete it.
This is not about writing new features. This is maintenance work, the kind that accumulates in every codebase. The bugs that nobody has time to fix. The cleanup that gets deferred. The report describes this type of work as “slow and painstaking,” noting that humans struggle with the volume of unfamiliar context required. Claude does not have that limitation.
The gap shows up in other ways. Anthropic now uses an automated Claude reviewer to check every proposed change to its codebase. When they ran a retrospective analysis, they found that this automated review would have caught roughly one-third of the bugs behind past incidents on Claude.ai before those bugs ever reached production. The engineers who wrote that original code are described as “among the best in the world.” Claude is catching mistakes that those engineers missed.
On code quality, the report states that Claude-written code was still somewhat worse than human-written code in late 2025. It has now reached rough parity. The agency expects it to be better within the year.
The implications go beyond Anthropic. If Claude can maintain an entire codebase at a rate that would take a human team years, what happens to the structure of software development across the industry? The bottleneck shifts. It is no longer about writing code faster. It is about deciding what code should be written at all.
The 52x Researcher
The report documents Claude’s performance on a specific internal benchmark that mirrors the research process itself. Each time Anthropic releases a new model, they give Claude a training script for a small AI system and ask it to make the code run as fast as possible while passing correctness checks. The goal is fixed. The method is not.
In May 2025, Claude Opus 4 averaged roughly a 3x speedup over the starting code. By April 2026, Claude Mythos Preview was achieving roughly 52x. For calibration: a skilled human researcher needs four to eight hours to reach a 4x speedup. Claude is operating at a level that is not just faster. It is operating in a different category entirely.
That number is not abstract. In April 2026, Anthropic published a demonstration of Claude running an entire open-ended research project from start to finish. The task: determine whether a weaker model can reliably supervise a stronger one. Claude proposed hypotheses, designed experiments, ran tests, shared findings with parallel agent instances, and iterated. Two human researchers spent about a week on the same problem and recovered roughly 23% of the performance gap between floor and ceiling. The agents recovered 97%, operating over 800 cumulative hours on roughly $18,000 in compute.
The caveats matter. The result did not transfer cleanly to production-scale models. Humans still chose the problem and defined the scoring rubric. But within those boundaries, the agents designed every experiment themselves. Direction setting was the only meaningful role a human played.
The report frames this as a narrowing gap. The things humans still do better at Anthropic are the high-level choices: which problems to work on, which results to trust, when an approach is a dead end. The report calls this “research taste.” But it also notes that Claude is improving here too. On a measure of next-step decision making during open-ended investigations, Anthropic’s best model in November 2025 made a better choice than the human researcher 51% of the time. By April 2026, that had grown to 64%.
The Three Futures
The report sketches three possible trajectories for what comes next.
The first is a stall. Progress slows because of bottlenecks in compute availability, safety constraints, or fundamental technical barriers that resist incremental improvement. In this scenario, recursive self-improvement remains a theoretical possibility rather than a practical reality.
The second is steady improvement. AI continues to get better, but at a pace that institutions and governments can roughly keep up with. Governance frameworks, safety research, and public understanding have time to evolve alongside the technology.
The third is takeoff. Recursive self-improvement accelerates. AI systems become capable enough to design and build their own successors without meaningful human direction. The loop closes. Progress compounds faster than any human institution can track.
The report does not predict which future will arrive. But it makes clear that the data points toward the third scenario as the one that requires the most preparation. The task horizon numbers alone are sobering. In March 2024, Claude Opus 3 could handle tasks that take a human about four minutes. In March 2025, Claude Sonnet 3.7 handled tasks of about 90 minutes. In March 2026, Claude Opus 4.6 was managing 12-hour tasks. METR, which measures these capabilities, found that Claude Mythos Preview could work for at least 16 hours and was at “the upper end of what METR can measure without new tasks.”
If this trend holds, the report notes, tasks that take a skilled person days could come into range this year. In 2027, AI systems could handle tasks that would take a person weeks.
Why Anthropic Is Asking to Slow Down
The call for a global pause is the report’s most significant political gesture. A frontier AI company publishing its own data showing dramatic acceleration and then using that data to argue for restraint is unusual. It puts Anthropic in a distinct position relative to its competitors.
The report’s framing suggests the agency sees the direction as clear even if the timing is uncertain. The human role is narrowing at every step. Once AI-authored code quality exceeds human code quality, the report notes, humans will stop writing code entirely and shift to reviewing only. But if they cannot review code as fast as Claude can generate it, human review becomes the bottleneck. The doing costs almost nothing in human time. The constraints shift to compute, safety, and governance.
The report also surfaces a quieter loss. One employee quoted in the document reflects on how work at Anthropic used to run on a gift economy of small favors between humans: requests for help running a script, each one creating a small debt and a small mutual awareness. Claude is faster. It creates no debt. But each of those interactions is a lost opportunity for human collaboration. The efficiency gain comes with a social cost that the report does not quantify but does not ignore either.
The argument for a pause is not a Luddite position. It is a structural one. The data in this report suggests that AI development is moving toward a point where human judgment is the only remaining input. The question is whether that judgment is being exercised thoughtfully enough, and whether the institutions that need to prepare for the third future are moving at the necessary speed.
The report does not say they are.