Anthropic: Claude writes 80% of its code, warns of recursive self-improvement

Anthropic revealed in an official statement that its AI assistant Claude is now responsible for writing more than 80% of the company's production code. Citing internal data, the firm said its engineers are merging eight times more code per day than they did in 2024, describing the trend as a possible path to recursive self-improvement—where AI autonomously builds a more capable successor.

The disclosure accompanies new internal research on Anthropic's Claude Mythos Preview model. According to AI monitoring account TestingCatalog, the research showed the system could operate autonomously for "at least" 16 hours, a duration that sits at the upper limit of what the evaluation group METR can currently measure. Anthropic's official account stated that the acceleration is "happening faster than we thought."

Commentator Min Choi highlighted that the coding statistic is "actually much bigger" than a simple productivity metric, framing it as evidence that AI systems are beginning to bootstrap their own development. Anthropic's own social media team echoed this sentiment, noting that the implications "deserve greater attention" as the industry approaches systems capable of improving themselves with minimal human intervention.