Anthropic’s chief product officer, Mike Krieger, says that nearly all of the code written inside the company is now produced by its own AI system, Claude. According to Krieger, engineers at the company regularly ship pull requests of 2,000 to 3,000 lines in which every line was generated by the model.
The claim, made in early February 2026, follows a prediction that Anthropic chief executive Dario Amodei made roughly a year earlier, when he said AI would soon be writing 90 percent of code. At the time that forecast was widely regarded as aggressive. Krieger’s account suggests that, at least within Anthropic, the share has moved past it and is now effectively 100 percent.
From writing code to directing it
Krieger described a shift in what engineers actually do. Rather than writing code line by line, the job has moved toward overall design, setting direction, reviewing the model’s output and making strategic decisions about what to build. Anthropic’s own internal survey of how staff use Claude, which the company published in late 2025, reported self-estimated productivity gains of more than 50 percent. That figure comes from the company’s own employees and has not been independently measured.
The company also says Claude Code, its command-line coding agent, has become considerably more autonomous. Earlier versions needed a human to step in after a limited number of actions; the current tool can carry out far longer sequences of steps on its own. Krieger said some engineers at the company have not written a line of code by hand in months and have shipped complete projects through the model, leaving repetitive work to the AI and spending their time on judgment calls.
What it means beyond Anthropic
Anthropic is an unusual case: it builds the model, has unlimited access to it and employs people who are highly motivated to use it. Other large technology companies report much lower shares of AI-generated code, though the gap appears to be narrowing. Microsoft chief executive Satya Nadella has said publicly that a significant portion of code in some Microsoft repositories is now written by AI, and OpenAI engineers have made similar claims about their own work.
The human role has not disappeared. Writing effective prompts, designing systems, and reviewing and approving code before it ships remain the responsibility of engineers. Krieger’s claim concerns who types the code, not who decides what the software should do or whether it is correct.
Market reaction
Investors have taken the trend seriously, particularly in relation to IT services companies whose business model rests on billing for conventional software development. If a frontier AI lab can automate most of its own coding, the question for the rest of the industry is how quickly similar practices spread and what that does to demand for traditional programming labor.
For now, the evidence is largely self-reported by the companies that sell the tools. Independent studies of AI coding productivity have produced mixed results, with some finding large gains and others finding that experienced developers can be slowed down by reviewing AI output. Anthropic’s experience is one data point from an environment built to favor the technology, and whether it generalizes to ordinary software teams remains an open question.
