Anthropic Says Claude Now Writes Over 80% of Its Own Code
On June 4, Anthropic disclosed that, based on internal data, "Claude is accelerating AI development itself," publicly suggesting for the first time a path toward "recursive self-improvement"—where AI autonomously builds a more capable successor. The company acknowledged it is "happening faster than we thought."
The company's official account (@AnthropicAI) posted that internal data shows Claude accelerating AI development, which could lead to "a possible path to recursive self-improvement, or AI autonomously building a more capable successor." Simultaneously, its research arm, The Anthropic Institute, released a report detailing the evidence.
The figures in the report are concrete. As of May 2026, over 80% of the code merged into Anthropic's production codebase was generated by Claude—up sharply from the low single digits before the Claude Code research preview. The amount of code merged per engineer reportedly reached about eight times the 2024 level. Furthermore, the unreleased frontier model Claude Mythos Preview achieved a 52x speedup on a code optimization task—a level said to require four to eight hours for a skilled researcher. Compared with the roughly 3x of Claude Opus 4 in May 2025, this underscores a leap in just a year. A March 2026 survey of 130 employees found a median fourfold increase in output when using Mythos Preview.
Behind this lies a perception that the initiative in AI development is gradually shifting from humans to AI. The company frames a transition from chatbot assistance around 2023, to coding agents in 2025, and now to autonomous agents in 2026. External benchmarks such as METR also confirm that the length of tasks AI can autonomously complete (time horizon) tends to double roughly every four months, with saturation progressing on metrics like SWE-bench and CORE-Bench. OpenAI noted similar concerns in a December 2025 blog post, and wariness of an "AI accelerating AI" feedback loop is spreading across the industry.
Anthropic maintains a safety-focused stance but acknowledged that progress is faster than expected, deciding to go public to prompt societal discussion. Axios the same day published an interview with the company's Jack Clark, reporting his view that AI progress will accelerate further and that briefings for lawmakers are being planned.
The post drew attention, gathering over 3,000 likes and more than 240,000 impressions as of June 4. Figures like 80% code generation and the 52x speedup were cited and spread as "SITUATION DETECTED," while developers shared workflows such as "Claude is not something you prompt, but something you build self-prompting systems with." Anthropic's Dan Altman also added emphasis on the importance of internal measurement. Positive and neutral opinions dominated, though some pointed out that a "judgment gap" remains before full self-improvement.