Anthropic’s AI assistant Claude now writes more than 80% of the code merged into the company’s production systems, a dramatic shift from last year. The figure, reported by the company in May 2026, marks a jump from low single-digit percentages before February 2025. Anthropic attributes the surge to rapid improvements in Claude’s coding abilities and says the trend is happening faster than expected.
From single digits to dominance
In early 2025, Claude contributed only a sliver of the code that made it into Anthropic’s products. By late 2025, that number began climbing. Now, as of May 2026, the AI handles the overwhelming majority. The company’s engineers haven’t been sidelined — they’re producing far more. In the second quarter of 2026, the typical Anthropic engineer merged eight times as much code per day as they did in 2024. That productivity gain comes from leaning on Claude for generation and review.
The shift didn’t happen overnight. Anthropic traces it to a series of improvements in Claude’s reasoning and coding accuracy, culminating in the Mythos Preview model this year.
Speed gains measured in multiples
Internal benchmarks show just how much faster Claude has become at certain coding tasks. Claude Opus 4, released in May 2025, averaged a 3x speedup on a test where the AI trains a small model — a proxy for more complex software work. A skilled human needs four to eight hours to achieve a 4x speedup on that same test. By April 2026, the Mythos Preview model hit a 52x speedup on that benchmark. That’s not a typo. Fifty-two times faster than the baseline.
Anthropic also tracks how well Claude recovers from a wrong turn. When shown a researcher’s mistaken approach, Mythos Preview picked a better next step 64% of the time. That’s up from 51% for Opus 4.5 in November 2025. The improvement suggests the AI is getting better not just at writing code, but at deciding what to write next.
Code quality catches up
Speed isn’t everything. Anthropic says Claude-written code was “somewhat worse” than human-written code in late 2025. By mid-2026, it has reached rough parity. The company expects it to be strictly better within a year. That timeline puts AI-generated code on track to surpass human output on reliability, efficiency, and clarity — at least by Anthropic’s internal measures.
The company frames this trajectory as a possible path to recursive self-improvement, where Claude builds its own successor. But it also cautions that Claude hasn’t yet shown the research judgment needed to choose which problems matter most. An AI that writes perfect code for the wrong task isn’t an improvement.
IPO filing and the bigger picture
Anthropic recently submitted a confidential IPO registration. The filing is still sealed, but the timing aligns with the company’s claims of accelerating progress. In its public statements, Anthropic says the trend — more code, faster, better — deserves greater attention from the industry. They’re not predicting what comes next.
What’s unresolved is whether Claude’s lack of research judgment will limit its trajectory or whether the next round of improvement will close that gap too. Anthropic has not said when the IPO might go public or what its valuation target is.
