There’s a sentence that sounds like science fiction but isn’t: the AI is writing most of its own code now. As of May 2026, over 80% of the code merged into Anthropic’s production codebase was authored by Claude Code, the company’s AI coding agent. Before its research preview launched in February 2025, AI-authored code at Anthropic sat in the low single digits.
The numbers behind the shift
Anthropic engineers are now merging roughly eight times as much code per day compared to 2024 levels. On internal optimization benchmarks, Claude achieved approximately 52x speedups by May 2026. For context, the same benchmarks showed roughly 3x speedups for skilled human programmers back in May 2025. In the span of twelve months, the AI went from modestly outpacing humans to lapping them dozens of times over.
One Anthropic engineer reportedly hasn’t written any code personally for about five months. Their job still exists, but it looks nothing like what “software engineer” meant two years ago. The role has shifted almost entirely toward direction, review, and architecture.
What the engineers actually do now
Anthropic is careful to frame this as collaboration rather than replacement. Humans still set strategic goals, make judgment calls on architecture, and review the code Claude produces before it ships. The company positions human oversight as the primary bottleneck on how fast new AI systems can be developed.
The question this raises is uncomfortable but unavoidable: if the AI writes the code that improves the AI that writes more code, where exactly does the feedback loop stabilize? Anthropic has publicly raised the need for verifiable global mechanisms that could potentially pause the development of frontier AI technology if circumstances warrant it.
Why this matters beyond Anthropic
There’s a meaningful difference between AI that assists with code completion and AI that authors 80% of a production codebase at one of the most prominent AI labs on the planet.
For investors watching the AI space, the speed of improvement is notable: going from 3x to 52x on internal benchmarks in a single year suggests the capability curve is still steepening, not flattening. Anthropic’s own call for global pause mechanisms hints that the industry may be approaching thresholds where self-regulation gives way to external oversight.