The AI Code Share Tracker: Every Verified Public Number on AI-Written Code, From Nadella's 'Maybe 20 to 30 Percent' to One Anthropic Engineer's 'Pretty Much 100%'

Public answers to how much code AI now writes at major technology companies run from 20% to 100%, and roughly four fifths of that spread is who is talking and what they are counting. Company-official figures cluster between about 65% and 90-plus. Every number outside that band comes from a narrower denominator or a louder speaker.

This ledger collects every dated public disclosure we could verify, tagged with the exact scope the speaker claimed. The anchors: Google says 75% of its new code is AI-generated as of April 2026 1. Anthropic's own spokesperson said 70% to 90% of code company-wide as of January 2026, the same month the head of its Claude Code team, Boris Cherny, claimed "pretty much 100%" 2; by May 2026 the company's CFO, Krishna Rao, put the figure at "90 plus percent of our code" 3. Microsoft CEO Satya Nadella's figure, 20% to 30% of code in some projects, dates to April 2025, with no updated figure on record as of this October 2026 refresh 4. And the newest official number on record is not a share at all but a volume: Anthropic's engineering blog said in October 2026 that merged code changes per day grew from about 2,500 in May to over 30,000 by September, growth the post attributes to Claude's output 5.

The ledger: who claimed what, when, and on which denominator

• Google (of new code, company-official): about 25% as of October 2024, 50% by fall 2025, and 75% by April 22 2026, with human engineers reviewing the AI output 1.
• Anthropic (all company code, spokesperson): "between 70% and 90%" as of January 2026; separately, about 90% of Claude Code's own code is written by Claude Code 2.
• Anthropic (company code, CFO): CFO Krishna Rao, on Patrick O'Shaughnessy's Invest Like the Best podcast in an episode published May 13 2026: "90 plus percent of our code is actually written by Claude Code" 3 6.
• Anthropic (all company code, Boris Cherny, head of Claude Code): "pretty much 100%" company-wide in January 2026, and 100% of his own code for more than two months, including 22 AI-written pull requests shipped in a single day 2.
• Anthropic (volume, engineering blog, not a share): code changes merged to the company monorepo averaged around 2,500 per day in May 2026 and topped 30,000 per day by September; the post attributes the growth to Claude's output and says the bottleneck has shifted from writing code to developer infrastructure built for human scale 5.
• OpenAI (one engineer's own code): roon, a pseudonymous account on X written by an OpenAI researcher, answered "100%, I don't write code anymore" in January 2026 2.
• OpenAI (president's stage claim, scope ambiguous): OpenAI president Greg Brockman, who cofounded the company in 2015, said at Sequoia's AI Ascent 2026 conference, in a talk uploaded April 30 2026: "If you look even over the course of December, we went from these agentic coding tools writing 20% of your code to writing 80% of your code," a change he dated to a single month. TNW reads the remark as roughly 80% of the company's code; the quote's "your code" leaves the denominator ambiguous, and TNW flags a second reading, lines written versus coding work AI touches at all. Brockman also said OpenAI ensures a human is responsible for all code that is merged 7 8.
• Microsoft (code in some projects, CEO): Nadella at Meta's LlamaCon, April 29 2025: "maybe 20 to 30 percent" of the code inside Microsoft repos today, in some projects 4 9. Fortune still cited this as the standing number in January 2026 2, and TechSpot still cited it as the standing number in May 2026 6.
• Snap (of new code, company statement): at least 65% of new code AI-generated under its new operating model, April 2026 1.
• Grindr (company code, CEO claim, unaudited): CEO George Arison said about 80% of the gay dating app's code is AI-written, produced by roughly 95 technical staff at 2.5x productivity, August 2026. No Grindr filing we reviewed carries the figure 10 11.
• Meta (internal targets, not achieved shares): a goal in place by Q4 2025 that 55% of code changes from engineers in some orgs be "Agent-Assisted," and a first-half 2026 expectation that 65% of engineers in its creation org write over 75% of committed code with AI 1.
• Salesforce (via Fortune's summary): a figure similar to Microsoft's 30% 2.

The 100% claims meet the 70-to-90% fine print

The sharpest disagreement in the dataset sits inside one company. In January 2026, Cherny put Anthropic's company-wide share at "pretty much 100%"; Anthropic's own spokesperson put it at 70% to 90% 2. That is a 10-to-30-point haircut between the claim that travels and the number the company stands behind, issued in the same month by the same company. Four months later the company line followed him most of the way: CFO Krishna Rao's "90 plus percent of our code" 3 leaves that haircut at ten points or less and is the only company-official figure in this ledger above 90%. Both 100% numbers still belong to individual engineers describing their own output. Grindr's 80%, unaudited and absent from the filings we reviewed, sits level with Anthropic's January official midpoint and ten points under the CFO's May floor. The pattern across the ledger: the louder the number, the smaller its denominator.

Two computed metrics: the AI Code Share Gap and Disclosure Velocity

AI Code Share Gap. Start with the loudest endpoints: 100% against Microsoft's 30% ceiling, a 70-point spread. Swap in official figures and it compresses. The frontier labs' disclosed figures are 90-plus for Anthropic (the CFO's stated floor, May 2026), 75% for Google, and 80% for OpenAI (Brockman's stated number, scope ambiguous), a median of 80; the median elsewhere, across Microsoft's 30, Snap's 65, and Grindr's 80, is 65. That is a 15-point gap. About 55 of the 70 headline points, roughly 78.6%, is speaker and scope rather than measured difference between companies. One caveat is structural: "new code," all code, and "some projects" are different denominators, so the compression is the finding, not the decimals.

Disclosure Velocity. Google is the only company with three dated official figures, and its pace doubled: the first 25 points took roughly a year, October 2024 to fall 2025; the next 25 took roughly six months, fall 2025 to April 2026 1. Anthropic's official line now holds two dated points, 70% to 90% in January and "90 plus percent" by mid-May 3; its newest official number, from the engineering blog on October 13, quotes volume instead of a share, and the volume moved harder than any share in this ledger has: merged changes per day went from around 2,500 in May to over 30,000 by September, at least a 12x rise in four months (our calculation from the cited figures) 5. Microsoft sits at the opposite pole: zero movement in the 18 months since April 2025, with no updated figure on record 4 2 6.

Why no single number will settle this

The denominators are different measurements. "Of new code" (Google, Snap) counts fresh output only. "Inside our repos today, in some of our projects" (Nadella) counts an unnamed subset. "Company-wide" (Anthropic) counts everything. "My own code" (Cherny, roon) counts one person. Adoption is a different metric entirely: Block (formerly Square) reports that approximately 95% of its engineers regularly use AI to assist development, which measures who uses the tools, not what share of code they produced 12. Industry benchmarks from Larridin's developer-productivity hub put average AI-assisted lines at 15% to 25% and top-quartile teams at 40% to 60% 12; every headline figure in this ledger tops that average band, though Microsoft's range is the closest call, since its stated 20% to 30% overlaps the 15-to-25 average and only its 30% ceiling clears it. That is what self-selection looks like. The population anchor sits lower still: a study of GitHub Python functions found about 29% of them AI-written in the U.S. 2 13.

Method: rows are English-language public statements attributed to a company, executive, or named or pseudonymous engineer, each with a dated source. Internal targets (Meta) and predictions, like Anthropic CEO Dario Amodei's January 2026 forecast that AI would handle most or all software engineering within six to twelve months 2, are listed but excluded from the computed medians. The Anthropic volume disclosure is likewise listed for its dates but contributes no share to the medians.