Claude Is Rapidly Building Its Own AI Successor, Anthropic Wants To Hit Pause
That trend, taken far enough and given enough compute, points toward what's called recursive self-improvement, an AI system capable of fully autonomously designing and developing its own successor. Anthropic says we're not there yet, and that it isn't inevitable. But it could happen sooner than most establishments are prepared for.
The internal numbers back that up. As of May 2026, more than 80% of code merged into Anthropic's codebase was authored by Claude. Before Claude Code launched in research preview in February 2025, that figure was in the low single digits. By the second quarter of 2026, the typical Anthropic engineer was merging eight times as much code per day as they were in 2024, though the company acknowledges that lines of code is an imperfect yardstick and almost certainly overstates the true productivity gain. Anthropic feels the underlying shift is real either way, which is much of the writing is now done by Claude, with engineers directing and reviewing rather than typing.
That increase in productivity lends itself to a broader capability curve, according to the company. The length of complex tasks AI models can reliably complete on their own has been doubling roughly every four months. Claude Opus 3 could handle software tasks taking a human about four minutes back in early 2024. Claude Sonnet 3.7 managed 90-minute tasks a year later. Claude Opus 4.6 has pushed that threshold to 12 hours. On SWE-bench, the standard real-world coding benchmark, models went from scoring in the low single digits to saturating the test in two years.
On a recurring internal optimization test, Claude Opus 4 hit roughly a 3x speedup in May 2025. By April 2026, Mythos Preview, Anthropic's still-unreleased frontier model, reached about 52x. A skilled human researcher needs four to eight hours to hit 4x. Mythos Preview is also the engine behind Project Glasswing, Anthropic's cybersecurity initiative that has already surfaced more than 10,000 high- and critical-severity vulnerabilities across major operating systems and software, a scale of discovery that would take traditional security teams years.
The one gap that keeps humans in the loop is what Anthropic calls research taste, the judgment to decide which problems are worth solving in the first place. Claude can execute a well-specified experiment or optimize code with superhuman speed, but large performance gaps persist when it comes to exercising judgment in choosing goals. That's the gap between AI today and a system that could autonomously design its own successor.
Some already feel human review itself is becoming a bottleneck. Anthropic now uses an automated Claude reviewer on all proposed code changes, and found in a retrospective analysis that it would have caught roughly a third of the bugs behind past incidents on claude.ai before they ever reached production, bugs that were missed by some of the best engineers in the world at building these systems. Claude is now capable of controlling your computer directly to handle tasks, and has been expanding into creative and enterprise software at a rapid rate. The autonomous footprint keeps growing.
On the policy front, Anthropic is calling for something that sounds almost counterproductive coming from a company racing to be the leader in AI. The company wants the world to have the option to slow, or temporarily pause, frontier AI development to let safety research and societal structures catch up. But it acknowledges a unilateral slowdown would simply change who the front-runner is without creating the wider calculated process that is currently missing. The goal is a global, verifiable mechanism where if the industry hits the brakes, everyone stops together.
Whether that kind of coordination is politically achievable is an open question, and most likely not realistic. But the company openly publishing data showing its own AI is already writing most of its code, and simultaneously calling for a brake on the technology, is about as candid a signal as one is likely to get from anyone inside the breakneck AI race.
