Anthropic published research on June 4, disclosing that its Claude AI now authors more than 80% of the code in its own codebase and warning that the technology may be approaching a threshold where it can autonomously design its own successors, calling on major AI powers to build a coordinated mechanism to slow or pause frontier development before that point arrives.
The report, When AI Builds Itself, was co-authored by Anthropic Institute head Marina Favaro and co-founder Jack Clark and marks the first major publication from the institute, which launched in March 2026.
The Numbers
As of May 2026, Claude wrote more than 80% of the code merged into Anthropic’s production codebase — up from low single digits before Claude Code launched in research preview in February 2025. Engineers now merge roughly eight times as much code per day as in 2024. In a March 2026 poll of 130 research-team employees, the median respondent estimated a 4x productivity gain with Mythos Preview, though Anthropic noted the true figure is likely somewhat lower.
In April 2026, Claude shipped more than 800 fixes that reduced a class of API errors by a factor of 1,000; the engineer overseeing Claude estimated a human would have taken four years. On research navigation — the harder test of whether AI can exercise judgment rather than just execute instructions — Mythos Preview outperformed skilled Anthropic researchers 64% of the time, up from 51% for Claude Opus 4.5 in November 2025. The report notes these moments were deliberately selected for having room to improve on the human’s choice.
Outside the company, AI task length has doubled every four months, down from every seven months a year earlier. Claude Opus 3 handled four-minute tasks in March 2024; Claude Opus 4.6 handles 12-hour tasks today. METR found Mythos Preview can sustain autonomous work for at least 16 hours — the upper limit of what it can currently measure.
The Warning
“Taken far enough, and given enough compute, that trend points to an AI system capable of fully autonomously designing and developing its own successor,” the report states. “We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for.”
AI can now match or outperform humans at executing well-specified research tasks but has not yet demonstrated the ability to choose which problems are worth working on — the gap between today’s systems and full autonomous self-improvement.
The Proposal
Despite headlines framing the report as a call for a “global freeze,” Anthropic’s actual proposal is conditional. A worldwide slowdown “would likely be a good thing,” the company argues — but only if multiple well-resourced labs in multiple countries agree to stop simultaneously under externally verifiable conditions. A unilateral halt, the report warns, would cede competitive and geopolitical ground without reducing systemic risk.
“Without a global coordination mechanism, companies and governments will have to make difficult decisions about safety while under competitive and geopolitical pressures,” the report states. Clark reportedly told BBC Newsnight: “You want the option to be able to take your foot off the gas and put your foot on the brake.”
Anthropic published the report one week after the company had confidentially filed for an initial public offering targeting a fall 2026 debut.
Read: Anthropic Files Confidential S-1 as AI IPO Wave Nears $3 Trillion
Information for this story was found via the sources and companies mentioned. The author has no securities or affiliations related to the organizations discussed. Not a recommendation to buy or sell. Always do additional research and consult a professional before purchasing a security. The author holds no licenses.
