Detailed Analysis
Anthropic has reached a remarkable internal milestone: the vast majority of code written at the company is now generated by its own AI system, Claude. Senior engineer Boris Cherny publicly disclosed that he has written 100% of his code using Claude Code and Opus 4.5 for over two months, with no manual edits whatsoever — a claim substantiated by his output of 22 and 27 pull requests on consecutive days, each entirely AI-written. While Cherny asserted that "pretty much 100%" of code across Anthropic is AI-generated, an official company spokesperson offered a more measured estimate of 70% to 90% company-wide. Most strikingly, Claude Code's own codebase is approximately 90% written by Claude Code itself — a recursive development loop that underscores just how deeply the tool has been integrated into Anthropic's engineering workflows.
Claude Code functions as an agentic coding assistant operating directly in the terminal, with access to a full computing environment that allows it to plan, edit, debug, and manage projects autonomously across multiple files. Rather than acting as a simple autocomplete tool, it accepts plain-language descriptions of desired outcomes and executes the full chain of changes required to implement them, including tracing bugs to their root causes and automating code review. The productivity implications are significant: Anthropic reports that technical onboarding, which previously took weeks, now typically requires only a few days, with engineers becoming productive almost immediately upon joining. Broader research from Anthropic also suggests AI assistance can accelerate certain coding tasks by up to 80%.
The Reddit thread where this development surfaced reflects a more ambivalent reception outside Anthropic's walls. The original poster raises pointed concerns about API quota consumption accelerating at unsustainable rates and warns of deeper structural risks — specifically, that rapid AI adoption is eroding engineering culture faster than the industry can adapt. The post urges engineers to advocate to management for preserving human-led development practices, framing the issue not merely as a quality concern but a financial one, noting that AI companies are becoming increasingly profit-driven as demand scales. This tension between raw productivity gains and the longer-term health of engineering ecosystems is a recurring friction point in the broader industry conversation.
The developments at Anthropic sit within a wider pattern of AI labs deploying their own models to accelerate internal development — a strategy sometimes called "dogfooding" taken to an extreme degree. Both Anthropic and OpenAI have reported converging on near-total AI code generation internally, signaling that the most aggressive adopters of AI coding tools are, paradoxically, the organizations building those tools. This creates a feedback loop: AI-assisted development speeds up the release of more capable AI coding tools, which in turn are immediately redeployed internally. The competitive pressure this generates across the broader software industry is substantial, as companies that once measured developer output in pull requests per sprint now contend with peers measuring it in dozens per day.
What remains unresolved — and what the Reddit commentary gestures toward — is the question of sustainability and quality at scale. High pull request velocity does not inherently map to well-architected, maintainable systems, and the long-term effects of reducing human editorial judgment in codebases are not yet well understood. The financial model also raises questions: as AI-generated code proliferates, the cost of API calls and compute becomes a meaningful operational variable for engineering teams, potentially concentrating AI-augmented productivity gains among better-resourced organizations. Whether the industry normalizes around hybrid human-AI development or continues accelerating toward near-full AI authorship will likely depend on how these cost, quality, and cultural pressures resolve over the next several years.
Read original article →