The Creator of Claude Code on The Hottest Piece of Software in the World

Boris Cherny, creator of Claude Code, says coding became a natural focus for Anthropic because models interact with the world through code, and that Sonnet 3.5 was what he calls the world's first really good coding model. Claude Code began as a side project, but Cherny argues it was in hindsight an obvious consequence of the safety mission, since real-world deployment is necessary to understand how models behave under actual conditions in ways that laboratory evaluations and mechanistic interpretability alone cannot provide.

Growth of Claude Code has been driven almost entirely by model improvements rather than changes to the product harness itself. Opus 4 launching in May, Opus 4.5 in November, and Opus 4.6 in February each represented inflection points. The product uses the same public Anthropic API available to all external customers, by design, as a dogfooding strategy. Cherny describes Claude Code as a significant contributor to Anthropic's business but says its primary stated purpose is learning about safety.

On security, prompt injection is the top concern raised by CISOs. Anthropic ran a competition giving external researchers one week and a twenty thousand dollar prize to successfully prompt inject Claude Code, and researchers were able to inject every model tested except Anthropic's. Mechanistic interpretability work produced neural probes that detect prompt injection in the model's neurons and can stop it in real time. Auto mode, which eliminates permission prompts, is described as safer than the prior prompt-based approach because much of the defense is now embedded in the model itself. Anthropic also runs Claude Code within an open-source sandbox restricting access to only explicitly granted files and websites, though sandbox breaches can occur and are actively monitored.

Code quality from the model improved to the point where Cherny says it is almost every time better than what he would have written himself, a shift he noticed definitively with the model referred to as Fable. He attributes much of this to the feedback loop allowing Claude Code to test its own output in a browser, iOS simulator, or running service, and says without that ability to iterate on its own output the model's code quality is significantly limited. Cherny says 100 percent of his personal code has been written by Claude Code since November of last year, and that all of Claude Code and all Anthropic products are written using it. Across Anthropic the average is approximately 90 percent of code written using Claude Code.

Cherny now runs anywhere from a few to sometimes hundreds or thousands of Claude instances collaborating simultaneously, and his role has shifted from deep-focus individual coding to directing those instances and generating tasks for them. He describes programming as having gone through two major abstraction leaps in two years after being roughly static for fifty years, first from autocomplete writing one line at a time to models writing entire files or features, and then to a level where the user talks to a model that talks to other models that write source code. He sees engineering roles segmenting into prototypers, builders, maintainers, scalers, and sweepers, and says scalers who take a product with product-market fit and grow it significantly are particularly sought after at Anthropic now.

The BUN team migrated their entire codebase from Zig to Rust in approximately eleven days using Claude Code with dynamic workflows at a cost of roughly one hundred and fifty thousand dollars in credits. Cherny estimates the same migration would previously have required several engineers working for roughly a year. Enterprise adoption follows a ladder model progressing from one Claude instance per engineer to ten, then one hundred, then one thousand instances per engineer at some companies today. Customers include Airbnb, Ramp, Salesforce, Deloitte, Accenture, NASA, large New York banks, and major pharma companies.

Cherny identifies what he calls a current product overhang, arguing that one-prompt-at-a-time interaction patterns are preventing users from experiencing the full intelligence of the model, and that loops, routines, and long-running agentic sessions are the way to unlock it. Claude Tag sessions have been running continuously for weeks at a time with coherent behavior, which he attributes to years of alignment and safety research rather than a recent development. On the question of whether humans lose foundational skills and deep craft knowledge as AI handles lower-order tasks, Cherny and the host leave the question unresolved, acknowledging that many intuit something meaningful is lost when no one retains the ability to do foundational work, but noting this remains empirically undetermined and is expected to become answerable within current lifetimes.