Anthropic Engineers Let Claude AI Build Its Own Next Versions — And the Results Are Staggering

There is a phrase circulating inside Anthropic that captures something genuinely new about how the company operates: AI doing 90% of the computer programming to build Anthropic's products — including its own AI. That is not a marketing line. It is a production metric. And the numbers flowing from it are rewriting what a software company can look like in 2026.

In late March, CEO Dario Amodei confirmed that from February to late March 2026 — a window of just 52 days — the Claude team shipped over 70 major features. The tools that anchored this sprint included Claude Code, the agentic coding assistant now generating $2.5 billion in annualized revenue, and Cowork, the desktop AI agent built for knowledge workers that four engineers assembled in ten days, with most of the code written by Claude Code itself. The engine behind this velocity is the same product the company is selling: Claude, building the next version of Claude.

The feedback loop that changes everything

Anthropic's engineers now use Claude for approximately 60% of their work, up from 28% just a year ago. The company reports this drives roughly a 27% boost in individual output — conservative relative to the 50% productivity gains cited by some internal teams — and enables the team to ship between 60 and 100 internal releases per day. The implication is structural, not incremental. When the tool you are building is also the tool that helps you build it, and that tool is improving, you get a compounding acceleration that traditional software development cycles cannot match.

This is not simply engineers using AI to autocomplete code faster. Claude Code reasons about entire codebases, traces data flows across files, and produces changes that require judgment, not just pattern recognition. Approximately 90% of Claude Code's own code is now written by Claude Code itself, according to Anthropic engineers. Four percent of all public GitHub commits are now authored by Claude Code — a figure that doubled in a single month — with projections of 20% or more by year-end.

"AI is doing 90% of the computer programming to build Anthropic's products, including its own AI." — Dario Amodei, CEO, Anthropic

What humans still do

Amodei has been direct about where humans remain essential. Engineers provide oversight, strategic direction, and judgment calls that Claude cannot make autonomously. The decision about what to build, which safety properties to preserve, and how to prioritize competing tradeoffs remain firmly in human hands. Claude accelerates the execution; the vision and accountability are human.

This matters because Anthropic's safety-first culture is not incidental to the product — it is baked into its architecture. The company trains Claude against a Constitutional AI framework whose written principles have grown from 2,700 words in 2023 to 23,000 words today. Letting Claude build Claude faster does not mean letting Claude decide what Claude should be. That distinction is precisely the thing Anthropic's founders left OpenAI to protect.

Seventy features in fifty-two days: what actually shipped

The 70-feature sprint from February to late March 2026 was anchored by two flagship products. Claude Code — already the fastest product ramp in B2B software history, going from zero to $2.5 billion annualized revenue in nine months — received a wave of capability upgrades that expanded its reach beyond developers into enterprise security and operations. Claude Code Security, launched as a limited research preview for Enterprise and Team customers, uses Opus 4.6 to scan production codebases by reasoning about data flows the way a human security researcher would, not by matching against known-bad patterns. In testing, it found over 500 vulnerabilities in well-reviewed open-source projects — including a decades-old bug in the Ghostscript PDF engine that traditional static analysis tools had missed entirely. Its false-positive rate is below 5%, compared to the 30–60% rate typical of conventional scanners.

Claude Cowork, launched January 12, 2026, extended Claude's agentic capabilities to knowledge workers who do not live in terminals. Four engineers built it in ten days using Claude Code. It creates and edits files directly on a user's machine, connects to Google Drive, Gmail, DocuSign, and FactSet through deep integrations, and handles workflows in legal, financial analysis, HR, engineering, and operations that previously required entire teams of analysts. Spotify reported up to a 90% reduction in engineering time on code migrations using Cowork, with over 650 AI-generated code changes shipped per month. At Novo Nordisk, Claude powers NovoScribe, which automates the regulatory documentation process for new medicines — one of the most time-intensive tasks in pharmaceutical development.

"I think that we are now transitioning almost into vibe working." — Scott White, Head of Product for Enterprise, Anthropic

The revenue number that has no precedent

In December 2024, Anthropic had $1 billion in annualized revenue. By March 2026 — fifteen months later — Bloomberg confirmed that figure had reached approximately $19 billion. SaaStr noted that no enterprise technology company in recorded history has compounded at this rate at this scale. Not Slack. Not Zoom. Not Snowflake. Not Salesforce in its early years. The growth rate has held above 10x annually for three consecutive years.

The breakdown explains the trajectory. Approximately 70–75% of Anthropic's revenue comes from API consumption — enterprises and developers paying per token. Around 80% of Anthropic's business is enterprise customers, a significantly higher enterprise mix than OpenAI. Over 500 customers now spend more than $1 million annually on Claude, up from a dozen two years ago. Eight of the Fortune 10 are Claude customers. One in five businesses on the Ramp payment platform now pays for Anthropic, up from one in 25 a year ago. And 79% of OpenAI's paying customers also pay for Anthropic — enterprises are not choosing between the two, they are running both for different workflows.

Claude Code is the single product driving the largest share of this momentum. It reached $2.5 billion in annualized revenue in nine months and more than doubled that figure since the start of 2026. Business subscriptions quadrupled in the first six weeks of the year. In February 2026, Anthropic raised $30 billion in a Series G round at a $380 billion valuation — the third most valuable private company on Earth. By late March, the company was weighing an IPO as early as October, with preliminary discussions underway with Goldman Sachs, JPMorgan, and Morgan Stanley targeting a raise of more than $60 billion.

The data leak that revealed what comes next

On March 27, 2026, a configuration error in Anthropic's content management system exposed nearly 3,000 unpublished assets — blog drafts, images, PDFs, and internal documents — to a publicly accessible data cache. Among them was a draft blog post announcing a new model called Claude Mythos.

The document described Mythos as "by far the most powerful AI model we've ever developed." It also referenced a model codenamed Capybara — described as a new tier above Opus, positioned as larger and more capable than Opus 4.6, with dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity. An Anthropic spokesperson confirmed that the company is "training and testing a new model" that represents "a step change" in AI performance and is "the most capable we've built to date," currently being trialed by early access customers.

The leaked materials also described Mythos's cybersecurity capabilities as surpassing all existing AI models by a significant margin — a claim that raises dual-use concerns Anthropic acknowledged directly. Given this, the company reportedly plans to release Mythos initially only to a select group of enterprise customers focused on cybersecurity defense, with no general public release planned in the near term.

The leak also revealed an upcoming invite-only retreat for European corporate CEOs at an 18th-century English country manor, where Amodei will demo unreleased Claude capabilities and discuss AI adoption with policymakers. Anthropic confirmed the event is part of an ongoing series of executive engagements the company has hosted over the past year.

The bigger picture

What Anthropic is demonstrating in 2026 is a proof of concept that the broader technology industry is watching with unease and fascination in equal measure. When the best AI coding tool is used to build a better AI coding tool, and that better tool is then used to build the next version, the rate of improvement does not add — it multiplies. The 70-feature sprint in 52 days is not a record to be celebrated and moved on from. It is a baseline that the next sprint will likely exceed.

Amodei's January 2026 essay, "The Adolescence of Technology," warned that AI will disrupt 50% of entry-level white-collar jobs over the next one to five years, and that AI more capable than any human in almost every domain may arrive as soon as 2027. The irony is not lost on observers that the company making that warning is also the one most visibly accelerating the timeline — by letting its AI build itself.