Anthropics Report on AI Selfimprovement "When AI Builds Itself"
Ronni Holmvig Strøm · 2026-06-05
The Anthropic Institute has published a report on its own progress toward recursive self-improvement, the point at which an AI system could autonomously design and train its successor. The headline the report invites is the obvious
The Anthropic Institute has published a report on its own progress toward recursive self-improvement, the point at which an AI system could autonomously design and train its successor. The headline the report invites is the obvious one: the loop is closing, the machine is learning to build the machine, and the timelines are short. Anthropic is careful to say the loop is not closed. "We are not there yet," they write, "and recursive self-improvement is not inevitable."
We want to read the report for something other than the countdown. Underneath the question of when AI builds AI sits a more immediate and better-evidenced finding, the kind of finding that rarely makes the headline because it is about people. The work of building a frontier model is being redivided between humans and machines, and the line is moving in a specific direction. The human is moving up a floor.
What the Outside Benchmarks Show
METR measures the task horizon of a model, the length of task a system can reliably complete on its own. That horizon has been doubling roughly every four months, up from a prior trend of every seven. In March 2024, Claude Opus 3 could handle software tasks that take a person about four minutes. A year later, Sonnet 3.7 managed tasks of about ninety minutes. A year after that, Opus 4.6 cleared twelve-hour tasks. Anthropic projects that days-long tasks come into range this year, and week-long tasks in 2027.
The coding and research benchmarks tell the same story from a different angle. SWE-bench hands a model a real open-source codebase and a real bug report and asks for a fix that passes the project's own tests. Scores went from low single digits to saturation in two years. CORE-Bench, which asks a model to reproduce a published paper's results, a precondition for conducting original research rather than merely assisting it, went from roughly 20% success in 2024 to saturation fifteen months later.
These are the curves that make people talk about closing the loop. They measure capability in the abstract. What they cannot measure is the thing the report is actually about: whether any of this is changing how AI itself gets built, and by whom.
What the Inside Data Shows
Here the report does something most capability writing does not. It opens Anthropic's own books.
As of May 2026, more than 80% of the code merged into Anthropic's codebase was authored by Claude. Before Claude Code entered research preview in February 2025, that figure sat in the low single digits. Lines of code merged per engineer per day stayed flat across the company's first four years, then turned upward in 2025 when Claude began running code rather than suggesting it, and steepened again in 2026 as the models started working autonomously over longer horizons. By the second quarter of 2026, the typical engineer was merging eight times as much code per day as in 2024.
The more interesting number is the one about open-ended work. Anthropic tracks how often its staff correct, redirect, or take over from Claude mid-task, and that rate has fallen steadily for a year, including on the hardest tier, the problems with no clear specification where the engineer does not yet know what the answer looks like. On that tier, Claude's success rate reached 76% in May 2026, up fifty percentage points in six months. The report gives a representative case: a routine upgrade began crashing tens of thousands of training jobs, an engineer handed Claude the live incident and cluster access and little else, and Claude worked through the running jobs one environment setting at a time, isolated a single obscure debugging flag, and confirmed a fix in about two hours. The normal estimate for that work was two to three days.
There is a quieter example that we find more telling than the throughput figures. In April 2026, Claude shipped more than 800 fixes that reduced a class of API errors by a factor of a thousand. The engineer overseeing the work estimated that a human would have needed four years, because solving other people's bugs is slow, unglamorous, and demands more unfamiliar context than a person comfortably holds.The work existed before. It simply never rose high enough on anyone's list to get done.