Anthropic reports 8x surge in code output, sees early signs of recursive self-improvement

Anthropic has observed an 8x increase in the amount of code merged into its internal codebase in 2026 compared to 2021–2024 trends, which the company interprets as early signs of what it calls prosaic recursive self-improvement — a compounding speedup in AI lab productivity driven by the very systems the lab is building.

The trend began in 2025 but accelerated sharply the following year, according to a post on the Anthropic Institute's website by Jack Clark, a co-founder of the company who spent recent months compiling the evidence internally. There are also early indications that as models grow more capable, they are handling increasingly difficult engineering and research tasks that previously fell to human staff.

Clark frames recursive self-improvement in two tiers. The maximalist version — an AI smart enough to design its own successor entirely — he estimates has a 60% chance of arriving by the end of 2028. The prosaic version, where AI simply makes the people building AI faster and more productive, appears to be underway already. But he's careful: the data so far is suggestive, not conclusive. The critical missing piece is whether AI systems can generate the kind of paradigm-shifting creative ideas that truly vault a field forward. That hasn't happened yet.

Anthropic is publishing this now because it believes the implications of recursive self-improvement — whether modest or maximal — are profound enough to warrant open discussion. "I cannot reconcile today's economy or society with a world where this technology continues to grow more powerful," Clark wrote, urging the research community to begin grappling with what self-accelerating AI development means for the future.