The 70–90% figure is real but deeply misleading without context. An official Anthropic spokesperson confirmed to Fortune in January 2026 that "company-wide the figure is between 70% and 90%," while specifying that about 90% of Claude Code's own codebase is self-written. However, the most rigorous internal analysis — by Anthropic employee Ryan Greenblatt in October 2025 — found the company-wide average for merged production code was closer to 50%, with only a minority of teams reaching 90%. The gap between these numbers reveals a definitional problem, not a factual one: the percentage changes dramatically depending on what counts as "code."
The claim originated as a forward-looking prediction from CEO Dario Amodei at the Council on Foreign Relations on March 10, 2025. His exact words: "I think we'll be there in three to six months — where AI is writing 90 percent of the code." This was about the industry broadly, not Anthropic specifically, and explicitly hedged as a prediction.
Six months later, at Dreamforce in mid-October 2025, Amodei retroactively declared his prediction fulfilled. Speaking with Salesforce CEO Marc Benioff, he said: "Within Anthropic and within a number of companies that we work with, that is absolutely true." When Benioff repeated the claim back — "90% of all code at Anthropic being written by the model today" — Amodei immediately qualified: "On many teams, you know, not uniformly." Most media coverage ran the "absolutely true" line and dropped the qualification.
Between these two moments, in approximately September 2025, Amodei used the notably vague formulation "70, 80, 90% of the code written at Anthropic is written by Claude," a range so broad it's almost unfalsifiable. This appears to be the genesis of the "70–90%" framing that later became the official corporate line.
The most important document in this story is a LessWrong post published October 22, 2025, by Ryan Greenblatt, an Anthropic-adjacent safety researcher with direct insider knowledge. Greenblatt systematically dismantled the 90% claim with three key findings. First, the company-wide average for AI-written lines of merged code was "much less than 90%, more like 50%." Second, some teams did reach ~90%, but they were a minority. Third, if you include all code that was "at all useful" — one-off scripts, throwaway experiments, code that never gets committed — the figure rises to closer to 90%.
Greenblatt's conclusion was blunt: "When Dario said 'this is absolutely true,' I don't agree: I don't think the prediction came true." Another Anthropic employee told him that while "90% of the code written by AI wasn't crazy," "most of the work happens in the remaining 10%." This distinction — between lines of code and engineering work — is the crux of why the statistic misleads.
Anthropic's own official research report, published December 2, 2025 and based on a survey of 132 engineers plus analysis of 200,000 Claude Code transcripts, conspicuously avoided citing a single code-generation percentage. Instead it reported that employees use Claude in 59% of their work, achieve a ~50% productivity boost, and can "fully delegate" only 0–20% of their tasks. The 67% increase in merged pull requests per engineer is real and impressive — but a long way from "AI writes all the code."
The claim has been stated by at least four distinct Anthropic sources, each with different scope and precision:
The pattern is unmistakable: numbers escalated over time (50% → 70–90% → "effectively 100%"), qualifications were progressively stripped, and product-specific figures were conflated with company-wide reality.
The entire claim hinges on a definitional ambiguity that most reporting ignores. There are at least four ways to measure "AI-written code," and they yield radically different answers:
| Metric | Approximate figure | Source |
|---|---|---|
| Lines of merged production code, company-wide average | ~50% | Greenblatt (Oct 2025) |
| All useful code including one-off scripts and experiments | Closer to 90% | Greenblatt (Oct 2025) |
| Lines of merged code on highest-adoption teams | ~90% | Greenblatt (Oct 2025) |
| Claude Code's own codebase | ~90% | Anthropic spokesperson (Jan 2026) |
| Individual power users (Boris Cherny) | 100% | Cherny (Jan 2026) |
| Official company-wide spokesperson figure | 70–90% | Anthropic spokesperson (Jan 2026) |
The jump from 50% to 90% comes from including throwaway code — scripts run once, test boilerplate, experimental prototypes — that inflates volume without reflecting proportional engineering effort. AI makes generating large quantities of low-complexity code nearly free, which pushes the percentage metric upward even when humans still do most of the hard design and debugging work.
The 70–90% figure is traceable to an official Anthropic spokesperson statement in Fortune (January 29, 2026), making it a legitimate, sourced claim. But it is not a neutral statistic. The most rigorous internal analysis puts merged production code at around 50% company-wide as of late 2025, with the higher end of the range achievable only by including ephemeral and low-stakes code. Anthropic's own formal research report avoided stating a single percentage entirely, instead presenting more nuanced metrics.
The claim has been distorted through secondary reporting in three consistent ways: Amodei's original prediction was retroactively treated as a fulfilled fact; his "on many teams, not uniformly" qualification was systematically dropped; and product-specific numbers (Claude Code writing itself) were conflated with company-wide averages. The figure is best understood as directionally accurate — AI is writing a very large share of code at Anthropic — but specifically unreliable as a precise statistic, because it depends entirely on what you count and who you ask. Anthropic has a clear incentive to promote the highest defensible number, and the escalation from 50% (insider estimate) to "effectively 100%" (CPO statement four months later) tracks more closely with promotional momentum than with a tripling of AI capability.