DEV Community: MEIKUIO

Who Owns Microsoft? What the Filings Actually Say

Egor Sivchik — Sat, 24 Oct 2026 14:34:41 +0000

Who owns Microsoft? No single person does. The largest holders are two index fund managers, Vanguard and BlackRock, and they are the only owners above five percent that Microsoft itself names in its October 2026 proxy statement. Everything else you have read about Gates and Ballmer runs on numbers filed in 2014 and 2019. Rechecked against SEC filings on October 23, 2026.

Who owns Microsoft? The short answer

Microsoft's only shareholders above five percent are Vanguard and BlackRock, according to its October 2026 proxy statement. The rest of the company sits with thousands of smaller funds, retail investors, and a handful of former executives who stopped reporting their stakes years ago. That last part matters more than any table you will see today, because most of the ownership figures circulating online have no date attached.

What the primary documents actually say, with the date each number was filed. Every row below comes from a specific SEC document, not from an aggregator.

Who Microsoft's filings name as owners, and when each number was filed
Microsoft proxy (DEF 14A) filed Oct 21, 2026, Vanguard Group 13G/A filed Mar 27, 2026, Vanguard Capital Management 13G filed Apr 30, 2026, Ballmer Form 4 filed Aug 20, 2014, Microsoft proxy filed Oct 16, 2019. All on SEC EDGAR, checked Oct 23, 2026

Holder	Shares	Stake	Document behind the number	When it was filed
Vanguard Capital Management LLC	555,988,240	7.48%	Schedule 13G filed Apr 30, 2026, quoted in Microsoft's Oct 21, 2026 proxy	Filed in 2026 - The proxy lists this Vanguard subsidiary, not The Vanguard Group, Inc. Its 13G reports sole voting power over 73,430,154 shares and sole dispositive power over 555,988,240.
The Vanguard Group, Inc.	0	0%	Schedule 13G/A filed Mar 27, 2026	Filed in 2026 - The filing gives the reason in its own words: "On January 12, 2026, The Vanguard Group, Inc. went through an internal realignment." This is the name most ranking pages still print next to 8 to 9 percent.
BlackRock, Inc.	540,020,228	7.30%	Schedule 13G/A filed Feb 13, 2024, quoted in Microsoft's Oct 21, 2026 proxy	Filed in 2024 - Microsoft's 2026 proxy says this line is "based on a Schedule 13G/A filed with the SEC on February 13, 2024": thirty-two months before the proxy itself.
Steve Ballmer	333,254,734	about 4.5%	His last Form 4, filed Aug 20, 2014	Filed before 2020 - No ownership filing on Microsoft under his name since. The percentage is our division of that 2014 count by the 7,426,426,719 shares outstanding on Sep 30, 2026 (4.49%), no filing states it.
Bill Gates	102,992,934	1.34%	Microsoft proxy filed Oct 16, 2019	Filed before 2020 - 1.34% is the figure that 2019 proxy printed. He left the board on March 13, 2020, and nothing under his name about Microsoft has been filed since. Microsoft's 2026 proxy does not mention him.
Satya Nadella	918,226	about 0.012%	Microsoft proxy filed Oct 21, 2026 (as of Sep 30, 2026)	Filed in 2026 - His most recent Form 4s were filed on September 1 and 2, 2026.
All 18 directors and executive officers	2,139,607	about 0.03%	Microsoft proxy filed Oct 21, 2026 (as of Sep 30, 2026)	Filed in 2026 - 2,139,607 of 7,426,426,719 shares is 0.029%. Every row in the proxy's insider table carries the footnote "Less than 1%".

Two rows carry the real story. Ballmer's percentage is our division of his 2014 count by the 7,426,426,719 shares outstanding on September 30, 2026, the record date in Microsoft's latest proxy, Gates's 1.34 percent is the figure his 2019 proxy printed. Their share counts are the last ones either man ever filed. Neither has been required to disclose anything since, and neither has. Keep that in mind every time a page tells you what they own "now."

The one document where Microsoft names its owners

Once a year, Microsoft files a proxy statement, a document called DEF 14A that invites shareholders to the annual meeting and, along the way, lists everyone who beneficially owns more than five percent of the stock. The current one was filed on October 21, 2026, for the meeting on December 8, 2026. It is the freshest document in which Microsoft itself says who owns it, and its ownership table contains exactly two names.

The table's own words: "This table lists all entities that are the beneficial owner of more than 5% of Microsoft common stock." Vanguard Capital Management LLC, 555,988,240 shares, 7.48 percent. BlackRock, 540,020,228 shares, 7.30 percent. That is the complete list. State Street, Fidelity, and T. Rowe Price, fixtures of every "top shareholders" article, are absent, because none of them clears the five percent bar that triggers a mention.

Two terms carry this whole subject and deserve one plain definition each. "Beneficial ownership" means having the power to vote or sell a share, whoever technically holds it, it is the SEC's way of asking who actually controls the stock rather than whose name sits on a certificate. And a proxy statement is written for shareholders about to vote, which is why it is the one place a company must tally, in one table, everyone whose stake could sway that vote. Companies file plenty of documents about money. The proxy is the one about power.

The same proxy also discloses what the people inside the building hold, and the numbers deserve more attention than they get. Satya Nadella: 918,226 shares. CFO Amy Hood: 459,226. Vice Chair Brad Smith: 372,454. Board members mostly hold four and five figure counts, with two directors, John Collison and Catherine MacGregor, at exactly zero. The group line reads "Directors and Executive Officers as a Group (18 people): 2,139,607," and every single row carries the same footnote symbol, an asterisk meaning "less than 1%." The eighteen people running one of the largest companies on earth together hold about three ten-thousandths of it.

Now read the footnotes nobody quotes. BlackRock's line is, in the proxy's own words, "based on a Schedule 13G/A filed with the SEC on February 13, 2024." The table is labeled "as of 09/30/2026," but that numerator comes from a filing made thirty-two months before the proxy itself. A 13G amendment is required only when a passive holder's position crosses certain thresholds, so Microsoft carries the last one forward. Vanguard's line changed this year for a different reason. The Vanguard Group, Inc., the name nearly every ranking page prints, filed a Schedule 13G/A on March 27, 2026 reporting zero Microsoft shares, explaining that "On January 12, 2026, The Vanguard Group, Inc. went through an internal realignment." Its subsidiary Vanguard Capital Management LLC filed its own Schedule 13G on April 30, 2026, for 555,988,240 shares, and that is the filing the new proxy quotes. So the company's official answer to "who owns Microsoft" mixes a 2026 filing with one from early 2024.

Microsoft's October 2026 proxy still sources BlackRock's stake to a filing from February 2024.

This is not sloppiness on Microsoft's part. It is how the disclosure system works, and it means the question "who owns Microsoft" has no single up-to-the-minute answer anywhere, including at Microsoft. What exists is a set of documents with dates, and the honest move is to read them with the dates attached. This habit predates 2025: the 2019 proxy sourced its Vanguard line to a 13G/A from February 2019 the same way.

The institutional layer, and why no two sites agree on it

Index funds hold Microsoft because it sits at or near the top of nearly every major index, and the money tracking those indexes has to hold it in proportion. Vanguard and BlackRock manage that money for tens of millions of fund investors. Neither firm picked Microsoft the way a stock picker would, and neither can dump it while their clients hold index funds. Their stakes mirror how much index money exists, and that keeps growing.

Quarterly, big managers file a Form 13F listing their US-listed holdings. One grounding note before the numbers: a 13F shows long positions in US-listed securities held by that specific filing entity, about 45 days after the quarter ends. It is the freshest public trace of institutional ownership, and also the reason the field disagrees with itself, because "that specific filing entity" stopped meaning what it used to.

Know what a 13F leaves out, too. It skips options and swaps, holdings in non-US vehicles, and anything sold after the quarter closed but before the filing landed. So even a fresh 13F describes a position that is at least six weeks old on the day you read it, held by one legal entity among possibly several. It is the best public instrument we have, and it still photographs only the past. Any page converting a 13F share count into a confident present-tense percentage has already made two silent assumptions on your behalf.

Whose money is that, really? When a page says "institutions own Microsoft," it is describing plumbing, not a cabal. Vanguard's stake is the pooled savings of people holding its index funds, BlackRock's iShares ETFs work the same way. The end owners are retirement accounts, pension plans, and ordinary brokerage customers, sliced into fund shares. This is also why the popular "institutional ownership" percentage swings so wildly across the field, from about 71 percent on one broker page to 81.55 on a tracker that cites GuruFocus: there is no single agreed definition of which holders to count as institutions, which entities to consolidate, and which quarter to freeze. The number is an editorial choice presented as a measurement.

Compare what pages ranking for this query print for one holder, Vanguard, as of October 23, 2026: 8.6 percent in the Microsoft forum answer that Google's AI Overview quotes, 8.7 on TMGM, 8.95 on an AI-funding tracker that copied last year's proxy, 9.44 on Investopedia, 9.49 on Capital.com, and 7.48 in Microsoft's own new proxy. Six values for one owner, and the newest filing behind any of them carries the smallest number. Part of that spread is timing. The rest happened quietly in 2026: Vanguard restructured how it files. After the realignment of January 12, 2026, The Vanguard Group, Inc. reported zero Microsoft shares on a 13G/A, a subsidiary filed a new 13G for 555,988,240, and in May and August 2026 the group filed 13F notices instead of full holdings reports, meaning other Vanguard entities now report its holdings. Aggregators that track one entity now print a fraction of the group total and call it "Vanguard." BlackRock's consolidated filer changed the same way: the entity that filed its reports through 2024 now sits under the name BlackRock Finance, Inc., with its last consolidated report covering mid-2024.

So when one aggregator prints 6.5 percent for a single Vanguard entity and a broker page says 9.4, both may be faithfully reading a real filing. They are just reading different slices of a company that no longer files as one piece, summing the separate Vanguard entities on the Q1 2026 aggregator data lands near 9.3 percent, a figure we pass along only with that warning attached. The practical takeaway for your own research: check which legal entity a number belongs to and what quarter it covers, before comparing it with anything.

Vanguard's stake, six ways
Ranking pages and MSFT proxy, checked Oct 23, 2026

Quarter / item	Vanguard stake as printed
MSFT proxy 2026	7.48
MS forum (AIO)	8.6
TMGM	8.7
AI-funding site	8.95
Investopedia	9.44
Capital.com	9.49

Who owns the most shares of Microsoft?

Vanguard holds the most Microsoft shares of any single owner: Microsoft's latest proxy, filed October 21, 2026, lists Vanguard Capital Management LLC at 555,988,240 shares or 7.48 percent, just ahead of BlackRock at 7.30. Among individuals, the last filed numbers put Steve Ballmer first at 333,254,734 shares, though that figure was filed in 2014 and nothing newer exists. No individual has ever appeared in the modern proxy's five percent table.

Sit for a second with how small the insiders' slice is. All eighteen of Microsoft's current directors and executive officers combined hold 2,139,607 shares, about 0.03 percent of the company. Popular data pages claim insiders own about 6 percent of Microsoft, WallStreetZen puts it at 6.25. That figure works only if you count Ballmer and Gates as insiders, which the SEC stopped doing when they left their board and officer roles, and if you use their 2014 and 2019 share counts as if they were current. The people who actually run Microsoft today own, collectively, a rounding error.

Does Bill Gates own Microsoft?

Bill Gates does not own Microsoft, and he has not run it for years. He last disclosed a stake in October 2019: 102,992,934 shares, 1.34 percent, in that year's proxy statement. He stepped down from the board on March 13, 2020, which Microsoft announced under the headline "Microsoft Co-Founder Bill Gates Steps Down from Board of Directors," and with that his obligation to report trades ended. Whatever he holds personally today is private information.

The confusion is understandable. He co-founded the company, held about 45 percent at the 1986 IPO by widely cited accounts, and spent three decades as its public face. But he has been giving shares away since the 1990s, mostly to the foundation that carries his name, and each ranking page models what remains differently. That is why you will find his "current" stake printed as 0.53 percent, 0.9, 1.34, and 1.39 on different pages of the same search results, none of them dated to a filing. The only dated number is the 2019 one.

A category error also hides inside "Gates's Microsoft wealth." His fortune has been managed for decades through Cascade Investment, a private office that spread it across railways, hotels, farmland, and dozens of listed companies. Forum threads asking how Gates keeps getting richer while selling Microsoft are asking about that machine, not about MSFT. By the time he left the board, Microsoft stock was one line in a much larger portfolio, which is exactly why his personal stake could quietly stop being public without anyone noticing for years.

The quiet exit: the foundation's last Microsoft share

Here is the part no ranking page has caught. The Bill & Melinda Gates Foundation Trust, the entity that invests the foundation's endowment and files a public 13F every quarter, reported zero Microsoft shares as of June 30, 2026. Not a reduced position. Zero, for the second quarter in a row.

The trust's own filings trace the exit. It held 28,957,247 Microsoft shares at the end of September 2024. A year later, after two heavy quarters of selling, the count was 9,191,207. By December 31, 2025 it stood at 7,691,207, worth about 3.7 billion dollars. Then the position vanishes: the report for March 31, 2026 lists no Microsoft at all, and the latest one, filed August 14, 2026 for the quarter ended June 30, lists 24 holdings, from Berkshire Hathaway to FedEx, with Microsoft on none of the lines.

The Gates Foundation Trust's Microsoft position, quarter by quarter
Bill & Melinda Gates Foundation Trust, Forms 13F-HR (CIK 0001166559), quarters ended Sep 30, 2024 to Jun 30, 2026, last filed Aug 14, 2026

Quarter end	Microsoft shares held	Reported value	Position
Sep 30, 2024	28,957,247	$12.46 billion	Large position - The starting point of this series: 28,957,247 shares.
Dec 31, 2024	28,457,247	$11.99 billion	Large position - 500,000 fewer shares than the quarter before.
Mar 31, 2025	28,457,247	$10.68 billion	Large position - Same share count as December, the lower value reflects Microsoft's share price.
Jun 30, 2025	26,191,207	$13.03 billion	Reduced - 2,266,040 fewer shares than the quarter before.
Sep 30, 2025	9,191,207	$4.76 billion	Reduced - 17,000,000 fewer shares in one quarter.
Dec 31, 2025	7,691,207	$3.72 billion	Reduced - 1,500,000 fewer shares, the last report that lists Microsoft at all.
Mar 31, 2026	0	$0	Zero shares - The report for this quarter lists no Microsoft position.
Jun 30, 2026	0	$0	Zero shares - Filed Aug 14, 2026: 24 holdings, none of them Microsoft. The next report, for Sep 30, 2026, is due by mid-November.

Source: Gates Foundation Trust Forms 13F-HR. Latest filing checked October 23, 2026. Chart originally made for Meikuio with Meiku Charts.

The mechanics behind the curve are ordinary, and that is exactly what makes the endpoint strange. Gates transfers shares to the foundation as gifts, the trust's investment office then sells them over time to fund grants, a pipeline that has moved tens of billions of dollars over two decades. Positions shrink, that is the design. But a pipeline shrinks toward some working balance, and this one went to zero and stayed there for two consecutive reports. Whether that is a permanent policy or a pause, the next 13F, due in mid-November 2026, will say, until then, zero is the entire public record. A 13F carries no explanations, and no filing obliges the trust to give one, so the why stays private.

Two boundaries keep this honest. A 13F covers only the trust's US-listed long positions, so it says nothing about Gates's personal shares, which stopped being public in 2019. And the trust exiting a stock is an endowment decision, made by its investment office, not a statement from Gates about Microsoft. What the filings do establish is narrower and stranger: the vehicle most people picture when they say "Gates's Microsoft billions" last reported a Microsoft position in the filing that covered December 2025, and both reports since show zero, and the pages ranking for this query still model his wealth as if it were there. The page that came closest, a license-key shop's blog, caught the September 2025 drawdown and then stopped updating before the zero became public in May 2026.

The Ballmer question: the number nobody dates

Steve Ballmer's last Microsoft ownership filing is a Form 4, the report insiders must file whenever their holdings change, filed August 20, 2014. It records a routine acquisition of 827 shares on August 19, bringing his total to 333,254,734. That figure, 333,254,734, is the exact share count WallStreetZen prints today, TMGM rounds it to 333.2 million, and the forum answer Google's AI Overview quotes gives him 4.5 percent, each presented as his current position, usually with a dollar value attached in 2026 prices.

We ran the census. A full-text search of SEC filings for "Ballmer" across ownership forms, Forms 3, 4, 13D and 13G, checked September 3, 2026, returns 34 documents, and the most recent one touching Microsoft is that August 2014 Form 4. A second pass on October 23, 2026 found nothing newer. His only later ownership filing of any kind covers a different company entirely. The reason is mechanical: he left the board in August 2014, ending insider reporting, and a stake below five percent triggers no 13G. A man estimated to hold four and a half percent of a 3.7 trillion dollar company has had no legal reason to tell anyone about it for over a decade, and he hasn't.

Every claim that Ballmer owns 4.5% of Microsoft today is a twelve-year-old number wearing today's date.

The arithmetic of his frozen number is worth thirty seconds, because it shows how precise the staleness is. After his last open-market sale in November 2010, his Form 4 showed 333,252,990 shares. Between 2010 and 2014 he collected small routine grants as a director, 827 shares in that final August 2014 filing alone, which is how the count landed on exactly 333,254,734. The pages printing that figure today are not quoting an estimate that happens to be old. They are quoting a specific twelve-year-old document, down to the last four digits, without naming it.

Be careful with what this does and does not prove. It does not mean Ballmer sold, and it does not mean he still holds every share, he has said publicly over the years that he kept his stock, and Forbes models his wealth on that basis, but a statement to a journalist is not a filing. The precise claim, the one that survives checking, is this: nobody outside Ballmer's family office knows his Microsoft position, and every site quoting 4.48 or 4.49 percent is dividing a twelve-year-old share count by this quarter's share total. The two camps arguing about whether he or Gates is "the largest individual shareholder" are both arguing from stale paper. The last filed numbers put Ballmer ahead, and that is the most anyone can honestly say.

Who founded Microsoft, and where did their shares go?

Bill Gates and Paul Allen founded Microsoft on April 4, 1975, in Albuquerque, New Mexico. The current proxy opens its shareholder letter with exactly that anniversary: "Fifty years ago, Bill Gates and Paul Allen started Microsoft with a simple but powerful idea." Gates's path from roughly 45 percent at the 1986 IPO to an undisclosed stake today ran through decades of scheduled selling and donation. Allen's ran out earlier: he stepped back from operations in 1983, his stake fell below the five percent disclosure line around 2000, and after that no filing tracked it. He died on October 15, 2018, and no public document discloses whether his estate still holds Microsoft at all. His affairs passed to his sister Jody Allen as executor, managed through the family office once called Vulcan, later renamed Vale Group. Somewhere in that structure the co-founder's remaining stake, if any survives, sits invisible to the public record, a quarter century after it last crossed a disclosure line.

That is the pattern worth noticing across all three founders' stories, Gates, Allen, and the early-employee generation: American disclosure rules follow roles and thresholds, not fame. The moment someone stops being a director, officer, or five percent holder, their ownership goes dark, no matter how large it once was. The public record of Microsoft's individual owners is a photo album, not a live feed.

The album does have dated pages, and they tell the dilution story cleanly. Roughly 45 percent at the March 13, 1986 IPO. About 24 percent a decade later, per accounts of the mid-1990s proxies. Down through years of pre-announced quarterly selling and foundation gifts to 1.34 percent in the 2019 proxy, the last page with his name on it. Each step is the same move: the company grew, new shares went to employees, and the founder converted paper into a philanthropy pipeline. Fifty years after Albuquerque, the founding stake has effectively left the building.

Who is the CEO of Microsoft, and how much does he own?

Satya Nadella is Microsoft's Chairman and Chief Executive Officer. He became CEO effective February 4, 2014, and Chairman of the Board in June 2021, both dates from Microsoft's own filings. Per the current proxy he beneficially owned 918,226 shares as of September 30, 2026, which is about 0.012 percent of the company, and fresh Form 4s show him selling scheduled tranches as recently as September 1, 2026. He runs Microsoft because the board employs him to, not because he owns it.

His incentive to care works through pay, not holdings. The proxy notes that over 95 percent of the CEO's annual target compensation is performance-based, most of it in stock awards that vest over years against growth and shareholder-return targets. So Nadella's wealth does track Microsoft's fate, just through a compensation contract the board designed rather than through a founder-sized stake. The regular Form 4 sales that headlines sometimes treat as ominous are the scheduled other half of that machine: awards vest, taxes come due, tranches get sold.

That distinction, control through role rather than ownership, is the actual answer to most versions of this question people type. Nobody who owns a large piece of Microsoft manages it, and nobody who manages it owns a large piece.

Above the CEO sits the board the shareholders elect. The 2026 proxy nominates twelve directors, eleven of them independent, with Sandra Peterson serving as Lead Independent Director, a role that exists precisely because Nadella chairs the board he reports to. Directors stand for election every year, and in an uncontested election a nominee needs more votes for than against. This is the machinery a seven percent shareholder actually touches: not a hotline to the CEO, but a ballot with twelve names on it, once a year, weighted by shares.

Control is votes, and every Microsoft share gets exactly one

Microsoft has one class of common stock, and every share carries exactly one vote. The company's own capital stock description spells it out: holders "are entitled to one vote per share on all matters voted on by the stockholders," with no cumulative voting, and the proxy summarizes it as a "single class of stock with equal voting rights." There are no founder super-shares, no golden share, and the 100 million authorized preferred shares sit unissued. Whoever holds the most stock holds the most votes, and today that is index money voting on behalf of fund investors.

This makes Microsoft almost boring next to the AI companies it funds. OpenAI's structure splits money from control so aggressively that a 26 percent foundation stake appoints the board while a larger investor gets none of it, and Anthropic routes ultimate control through a purpose trust that owns almost no equity. Microsoft is the control group in that experiment: the oldest, plainest answer to "who decides," which is one share, one vote, majority rules. A seven percent index manager can vote seven percent, and its stewardship teams, the departments that cast the funds' votes, mostly follow standard governance policy rather than any agenda, because more than ninety percent of that money passively tracks indexes. Together the two managers named in the proxy vote roughly a seventh of the company: enough to matter in a close proposal, nowhere near enough to run it.

One more mechanism quietly reshapes everyone's percentage: buybacks. Microsoft's board has authorized repeated repurchase programs, 40 billion dollar authorizations in 2016 and 2019, then 60 billion in September 2021 and again in September 2024. In fiscal 2026 the company repurchased roughly 36 million shares while issuing about 29 million in employee equity, so the total count fell from 7,434 million at the start of the year to 7,427 million at its end. When the denominator shrinks, every holder's percentage drifts up without anyone buying a share. Ballmer's frozen 2014 share count, for instance, computes to a slightly higher percentage today than it did the day he filed it.

You can watch the denominator move on the covers of three consecutive filings: 7,433,166,379 shares as of July 24, 2025, then 7,428,434,704 as of April 23, 2026, then 7,425,545,491 as of July 23, 2026. About seven and a half million shares evaporated in a year, at a company that also grants stock to over two hundred thousand employees. The new proxy's record-date count, 7,426,426,719 on September 30, 2026, sits about 880,000 above July's, so the shrink shows up across years rather than in every quarter. That slow shrink is the closest thing Microsoft's ownership has to a plot: fewer shares, the same index money, and percentages that creep upward for everyone left holding.

Insiders: about 6% or 0.03%?
Data pages and MSFT proxy, checked Oct 23, 2026

Quarter / item	Insider ownership as printed
WallStreetZen	6.25
TMGM	6
MSFT proxy 2026	0.03

What Microsoft owns: the flip side of the question

Half the people typing "who owns Microsoft" turn out to be asking the opposite question, judging by the related searches Google shows: who owns Xbox, who owns LinkedIn, who owned GitHub before Microsoft. The census below comes from Microsoft's own deal filings, announced value versus what the accounting later recorded where they differ.

The before-owners answer the nostalgia half of those searches. LinkedIn was an independent public company, co-founded by Reid Hoffman, until Microsoft paid 196 dollars a share for it in 2016, Hoffman sits on Microsoft's board today. GitHub was a venture-backed private company when the 7.5 billion dollar stock deal closed in October 2018. Activision Blizzard, the biggest deal in Microsoft's history and the one that made "who owns Call of Duty" a Microsoft question, traded publicly until the merger took effect on October 13, 2023 after nearly two years of regulatory fighting. All three now sit inside Microsoft's reporting segments as wholly owned pieces of one balance sheet.

What Microsoft owns, and one large thing it does not
Microsoft 10-Q and 10-K filings and deal announcements on SEC EDGAR, OpenAI interest per Microsoft's FY2026 10-K

Company	Deal closed	Price	Status today
Activision Blizzard (Xbox games)	Oct 13, 2023	$68.7 billion announced	Owned outright - The merger took effect on October 13, 2023, after nearly two years of regulatory review.
LinkedIn	Dec 8, 2016	$26.2 billion, $196 per share	Owned outright - Co-founder Reid Hoffman sits on Microsoft's board.
Nuance (healthcare AI)	Mar 4, 2022	$18.8 billion purchase price	Owned outright - Announced at $19.7 billion, $18.8 billion is the purchase price Microsoft's filings record.
GitHub	Oct 25, 2018	$7.5 billion in stock	Owned outright - Microsoft's 10-Q for the quarter ended December 31, 2018 gives the closing date as October 25, 2018.
ZeniMax / Bethesda	Mar 9, 2021	$8.1 billion purchase price	Owned outright - Announced at $7.5 billion, Microsoft's 10-Q records a total purchase price of $8.1 billion.
Mojang (Minecraft)	Late 2014	$2.5 billion announced	Owned outright - Announced September 15, 2014.
Skype	Oct 13, 2011	$8.5 billion announced	Retired - Microsoft retired Skype in 2025.
OpenAI	Investment, not a purchase	about 25% interest per Microsoft's 10-K	Minority stake - Microsoft's 10-K reports an interest of approximately 25 percent and does not consolidate OpenAI. Per Microsoft's earnings releases, the position swung from a $3.6 billion net loss in fiscal 2025 to a $5 billion net gain in fiscal 2026.

The OpenAI row answers the single most repeated misconception in forums: Microsoft does not own OpenAI. Its own 10-K describes an interest of approximately 25 percent, an economic share of results that comes with none of the board control ownership usually implies, the full mechanics, and why that percentage moves, live in our breakdown of who owns ChatGPT. The accounting shows how un-subsidiary the relationship is: per Microsoft's own earnings releases, the OpenAI position swung from a 3.6 billion dollar net loss in fiscal 2025 to a 5 billion dollar net gain in fiscal 2026, valuation weather passing through the income statement, which is what happens to investments and never to divisions. That same pricing fog surrounds private AI rounds across the market: see our Perplexity valuation reality check, where four conflicting numbers claim to measure the same one-year-old round. Copilot, by contrast, is Microsoft's own product through and through, built on licensed OpenAI models plus Microsoft's in-house MAI models, which it trains itself.

How Microsoft compares with the rest of Big Tech

Microsoft's plain structure is the exception among its peers, not the rule. Meta and Alphabet both run dual-class structures in which founders hold special high-vote shares, which is why Mark Zuckerberg can control Meta outright while owning a minority of its economics, and why Google's founders kept decisive votes long after selling most of their stake. At Microsoft that instrument simply does not exist. When Gates wanted to reduce his exposure, every share he sold surrendered exactly its share of the vote, and over four decades that arithmetic transferred control from a founder to an index-fund electorate one sale at a time.

Apple sits closest to Microsoft's model: single class, dispersed ownership, the same two index giants at the top of the register. That pattern, where the "owner" of a trillion-dollar company is effectively everyone with a retirement account, is quietly becoming the default state of mature tech. The founder-controlled structure makes headlines, the Microsoft structure holds most of the market's weight.

What the top results get wrong

We read every page ranking on the first page of Google for who owns Microsoft, on September 3 and again on October 23, 2026, and checked their numbers against the filings. The pattern is not that they are careless. It is that almost none of them date anything, and undated ownership numbers rot silently.

The specifics, as of our October 23, 2026 check. One stock data page, WallStreetZen, tells its readers Microsoft is "owned by 0.07% institutional shareholders, 6.25% Microsoft insiders, and 93.68% retail investors," an inverted split that contradicts both every other page and the table printed directly above it. The third organic result, another data aggregator, shows dashes where its institutional percentages should be. A broker's analysis on Capital.com names Bill Gates the largest individual shareholder without printing a single number for him, while TMGM and WallStreetZen crown Ballmer using his 2014 count. Gates's stake runs from 0.53 percent on TMGM to 1.39 on WallStreetZen. The institutional total swings from about 71 to about 82 percent depending on who is counting and what they count. And the answer engines inherit all of it. Google's AI Overview answers this query with "Former CEO Steve Ballmer is the largest individual shareholder (~4.5%)" and puts Vanguard at about 8.6 percent, citing a Microsoft forum reply that announces it was written by "Bing Chat running on ChatGPT4" on January 5, 2024. In September, the same overview cited a Microsoft license-key shop.

Google's AI Overview for who owns Microsoft now cites a 2024 chatbot reply.

What the top results say about Microsoft's owners, and what the filings say
Pages ranking on Google (US) for "who owns microsoft", read Oct 23, 2026, SEC EDGAR filings named in each row

Claim as printed	Range on the results pages	What the filings say	Verdict
Vanguard's stake	8.6% to 9.49%	7.48% for Vanguard Capital Management LLC in Microsoft's Oct 2026 proxy, The Vanguard Group itself reported zero on Mar 27, 2026	Partly right - None of the five figures on the results pages matches the holder and number in Microsoft's 2026 proxy, and only Investopedia and Capital.com say what date their number describes.
Gates's stake	0.53% to 1.39%, or "largest individual shareholder"	1.34% in Microsoft's October 2019 proxy, nothing filed under his name since	Contradicted by filings - The 2019 proxy is the last document that states his stake. Microsoft's 2026 proxy does not mention him.
Ballmer's stake	4.48% to 4.5%, presented as current	333,254,734 shares on his Aug 20, 2014 Form 4, nothing filed since	Contradicted by filings - The share count is real. Whether he holds it today is not public: he left the board in 2014 and has filed nothing on Microsoft since.
"Institutions own"	about 71% to 81.55%	No filing defines one total, it depends on which entities and which quarter are counted	Partly right - TMGM prints about 71%, an AI-funding site cites GuruFocus at 81.55% as of July 10, 2026, and one data aggregator in the top three shows dashes where its institutional figures should be.
"Insiders own"	about 6% to 6.25%	0.03% for all 18 current directors and executive officers (2,139,607 shares), Microsoft's Oct 2026 proxy	Contradicted by filings - The 6% figure works only if Ballmer and Gates are counted as insiders, as WallStreetZen's insider list does, with Ballmer's 2014 share count at the top.

Credit where due: Investopedia's page, updated January 7, 2026, dates what it prints, Vanguard's 9.44 percent and BlackRock's 7.96 percent "as of Sept. 30, 2025," and each executive's share count to a specific day. In September, a broker education page from Admiral Markets did the same for Ballmer and Gates, it no longer ranks on the first two pages. Everyone else launders old filings into 2026 by omission.

Watch how the error now compounds. An aggregator freezes a 2014 number. A blog copies the aggregator. Google's AI Overview summarizes the blog. A chatbot answer citing the same chain gets posted to a support forum, ranks, and becomes a source itself. That last step is no longer hypothetical: in October 2026, Google's AI Overview for this query cited exactly such a forum answer. At no point does anyone reopen the filing, so the loop can run for another decade unless the reader breaks it. Which brings us to the only durable fix.

How to check any ownership number yourself

Every number in this article came from EDGAR, the SEC's free public database at sec.gov, and checking one takes minutes once you know which of three documents to open. No account, no paywall, no data subscription.

For the official owner list: search EDGAR for Microsoft's latest DEF 14A and find the "Principal Shareholders" and "Stock Ownership Information" tables. This is the company's own answer, footnotes included. Read the footnotes, as this article shows, that is where the real dates live.
For a fund's current position: open the manager's most recent Form 13F and search the table for Microsoft or its identifier, CUSIP 594918104. Note the quarter-end date and which legal entity filed, because large managers now report through several.
For any individual: run their name through EDGAR's full-text search filtered to Forms 3, 4, and 13D/G. The newest hit is the freshest thing the public record holds, and if it is from 2014, then so is every "current" figure built on it.

The denominator for any percentage you compute lives on the cover page of the latest 10-K or 10-Q, updated to a date a few weeks before filing. Divide, and you will often find you have just produced a fresher number than anything on the first page of Google, which is a strange thing to be able to say about the most-watched company on earth.

Can you buy a stake in Microsoft?

Yes. Microsoft trades on NASDAQ under the ticker MSFT, and one share makes you an owner with the same one vote per share as Vanguard gets, which is the whole difference between this company and the AI startups around it. Ownership at that scale is mostly symbolic, but it is real: you can vote your shares at the December annual meeting, on directors, on executive pay, and on the shareholder proposals that reach the ballot each year. You cannot buy OpenAI, as our breakdown of every claimed path to OpenAI stock shows in detail, but the company holding roughly a quarter of it trades publicly all day. Whether MSFT belongs in a portfolio is a different question, and one this page deliberately does not answer, if you came here on the investing track, the sorting of which AI stocks actually book revenue and which quantum tickers run ahead of their physics is where we do that work.

Which answer did you actually need?

People arrive at this question from different places. The short routes, by reason for asking:

Settling an argument about Gates: he last disclosed 1.34 percent in 2019, left the board in 2020, and his foundation's trust now reports zero Microsoft shares. He neither owns nor runs it.
Wondering who controls the company: no one person. One share, one vote, a twelve-seat board, and index funds as the biggest voters. That is who owns Microsoft in the sense the question usually means.
Checking who profits when you buy Windows or Xbox: the shareholder base above, which by the largest measures is mostly index funds, meaning pension savers and 401(k) holders in aggregate.
Researching before buying the stock: the ownership table is public paper with dates, read the dates first, then the numbers, and treat anything undated as decoration.

Ownership figures move with every 13F season and every proxy. All numbers on this page were read directly from SEC filings on September 3, 2026 and rechecked against Microsoft's October 21, 2026 proxy on October 23, 2026, the next scheduled updates to the record are the November 2026 13F wave and Microsoft's next quarterly report.

The register: Microsoft's proxy (filed October 21, 2026) names exactly two owners above 5 percent, Vanguard Capital Management LLC 555,988,240 shares (7.48 percent, from a Schedule 13G filed April 30, 2026) and BlackRock 540,020,228 (7.30 percent, from a Schedule 13G/A filed February 13, 2024).
Denominator: 7,425,545,491 shares outstanding as of July 23, 2026, per the 10-K cover, and 7,426,426,719 on September 30, 2026, the proxy's record date, the count fell during fiscal 2026 as buybacks outran new grants.
Insiders: all 18 current directors and executive officers together hold 2,139,607 shares, about 0.03 percent, Nadella holds 918,226.
Frozen numbers: Ballmer's 333,254,734 dates to his last Form 4 of August 20, 2014, Gates's 102,992,934 (1.34 percent) to the October 2019 proxy, neither has filed since.
The exit: the Gates Foundation Trust reported zero Microsoft shares as of March 31 and June 30, 2026, down from 28,957,247 in September 2024, per its own quarterly 13F filings.
Structure: one class of common stock, one vote per share, no cumulative voting, preferred shares authorized but unissued.
Sources: Microsoft DEF 14A (October 21, 2026), Vanguard Schedule 13G/A (March 27, 2026) and Vanguard Capital Management Schedule 13G (April 30, 2026), Microsoft Form 10-K for fiscal 2026 (July 29, 2026), Gates Foundation Trust Forms 13F-HR (2024 through August 14, 2026), SEC EDGAR full-text search of Ballmer and Gates ownership filings, Microsoft 2019 DEF 14A, Microsoft board announcement of March 13, 2020.

Frequently Asked Questions

Nobody holds a majority. The largest owner Microsoft names is Vanguard Capital Management at 7.48 percent, per the October 2026 proxy, with BlackRock at 7.30. Institutional investors collectively hold most of the stock, but no single entity comes anywhere near 50 percent.

Not in any controlling sense. His last disclosed stake was 102,992,934 shares, 1.34 percent, in the 2019 proxy. He left the board in March 2020, ending his reporting duty, so his current personal holding is unknown.

By the last filed numbers, yes: Ballmer's 2014 filing shows 333,254,734 shares against Gates's 102,992,934 from 2019. Neither figure is current, and neither man has been required to file since, so today's true ranking is unknowable from public records.

Yes. Microsoft has traded publicly since its IPO on March 13, 1986, and lists on NASDAQ under the ticker MSFT. Anyone can buy shares, and each share carries one vote.

No. Microsoft's own 10-K describes an interest of approximately 25 percent in OpenAI, an investment that carries no board control. OpenAI remains governed by its foundation, Microsoft is its largest outside backer, not its owner.

Microsoft owns Copilot outright. It is a Microsoft product built on a mix of licensed OpenAI models and Microsoft's own MAI model family, so the brand, the product, and the infrastructure are all Microsoft's.

No. No government body appears among Microsoft's disclosed holders. The largest state-linked investor is Norway's sovereign wealth fund, whose reported holding as of June 30, 2026 works out to roughly 1.4 percent, and that is Norway's pension money, not a controlling government stake.

Zero, as of its latest filings. The Gates Foundation Trust reported no Microsoft shares for the quarters ending March 31 and June 30, 2026, after selling down from about 29 million shares in late 2024.

None. Microsoft has no parent company and has never had one since going public in 1986. It sits at the top of its own corporate tree, owning subsidiaries like LinkedIn, GitHub, and Activision Blizzard rather than being owned.

7,425,545,491 shares of common stock were outstanding as of July 23, 2026, per the cover of Microsoft's latest 10-K. The count shrinks slightly most years because buybacks retire more shares than employee stock programs create.

My Take

I went into this expecting to write a shareholder list and came out writing about dates. The thing that actually separates a usable answer from a recycled one here is a single column most pages delete: when the number was filed. Once you add it back, the confident consensus dissolves into a 2024 filing inside a 2026 proxy, a 2014 share count sold as current, and a foundation position that quietly hit zero while every wealth tracker kept modeling it.

If you take one habit from this page, make it this: before you repeat any ownership number, for Microsoft or anyone else, find its date. If a page will not tell you when a figure was filed, it is not reporting the number, it is decorating with it. The filings are free, they are public, and as this article shows, reading them beats trusting the first page of Google by an embarrassing margin.

Originally published on Meikuio — Who Owns Microsoft? What the Filings Actually Say.

ChatGPT vs Claude: Which Is Better for Writing?

Egor Sivchik — Wed, 21 Oct 2026 15:52:19 +0000

By Ryan Kowalski.

ChatGPT vs Claude for writing came down to a clear trade-off in these three samples. Claude used more concrete details and a stronger point of view. ChatGPT was steadier in tone and did not add the unsupported performance claim that appeared in Claude’s ad copy. I ran both through the same blog, marketing and structured-writing prompts. The results show where each draft needed an editor, rather than proving that either model is always the better writer.

Quick verdict

The twelve scores are reproduced from the June 3 comparison. They are subjective editorial ratings, with no published scoring rubric, not benchmark measurements of current models. Accuracy labels the original Factual reliability row. Made with Meiku Charts.

The short version: I preferred Claude for voice in this test, while ChatGPT’s marketing sample stayed closer to the supplied facts. Both followed the structured brief. Choose by the kind of editing your work needs, and check factual claims from either tool.

How I tested ChatGPT and Claude

I gave each tool the same three prompts, word for word. The comparison was published on June 3, 2026, using GPT-5.5 in ChatGPT and Sonnet 4.6 in Claude, in their desktop apps. These are the model labels recorded in that test, not a claim about the models available on every plan today. Both answered quickly enough that speed was not a deciding factor. The prompts and original examples below are preserved, with plan information checked separately on October 21, 2026.

Each prompt targeted a different writing job. Want to see the difference for yourself? Copy the three prompts below and run them on whatever plan you’re on, free or paid, then compare what comes back.

Prompt 1, tone and human feel: “Write a 150-word intro for a blog post titled ‘How to Stop Procrastinating When You Work From Home.’ Make it engaging, conversational, and avoid clichés. Don’t use phrases like ‘in today’s fast-paced world.'”

Prompt 2, marketing copy: “Write 3 short Instagram ad captions for a productivity app called ‘FocusFlow.’ Target audience: freelancers who struggle with distractions. Each caption max 2 sentences, include a call to action.”

Prompt 3, following instructions: “Write a 400-word section explaining the Pomodoro Technique. Use exactly 3 subheadings, include one numbered list of steps, and end with a single practical tip. Keep the tone practical, not academic.”

Blog posts and articles

This is where Claude pulled ahead first. Asked for a blog intro, ChatGPT opened with a competent, friendly line: “Working from home sounds like a dream until you realize how easy it is to drift through a morning.” Accurate, readable, and a little generic, the kind of opening you’ve read a hundred times.

Claude went for specifics. Its intro landed on “it’s 4pm and you’ve deep-cleaned the kitchen, watched three ‘educational’ YouTube videos, and reorganized your desktop icons. Again.” Ending on “Again” gave that opening a familiar, self-aware tone.

The structure differed too: ChatGPT delivered two smooth, even paragraphs, while Claude broke its intro into three shorter ones and leaned on one-line punches like “The couch is right there. Nobody’s watching.” ChatGPT’s version was solid but closed on a stock promise about what “you’ll learn in this guide.” Claude reframed procrastination as an environment-and-systems issue rather than a willpower problem. That gave this sample an angle instead of just a setup. For the blog opening in this test, I preferred Claude’s voice.

ChatGPT’s blog opening from the June comparison.

Claude’s blog opening from the same prompt in the June comparison.

Marketing copy

Here the two tools split, and it’s the most useful comparison in this test. Both wrote three Instagram captions for a fictional focus app aimed at freelancers. ChatGPT’s were clean and ready to publish, but the calls to action were the kind you forget instantly: “Try it today,” “Start now.”

Claude did something the prompt never asked for: it organized the three captions by strategy, pain-first, result-first, and direct. The language was sharper and tuned to the reader (“every ‘quick check’ is costing you money,” “without the guilt spiral”). For a marketer, that framing is useful.

But Claude also added a claim the brief did not support. One caption promised “3× more deep work hours” for an app that does not exist. I nearly dropped it into a brief before noticing that the number came from nowhere. A real campaign would need evidence for that claim, or the claim would need to go. ChatGPT’s sample did not add a numerical performance promise. That is an observation from this task, not a guarantee of factual accuracy in future drafts.

ChatGPT’s captions for the fictional FocusFlow app.

Claude’s captions include the unsupported “3×” claim discussed above.

Long-form and following instructions

The last prompt was deliberately rigid: a 400-word Pomodoro explainer with exactly three subheadings, one numbered list, and a single practical tip at the end. Both tools hit every requirement. No format slips, no missing pieces.

The difference was in approach, and you could see it in the section titles. ChatGPT labelled its three subheadings exactly as you’d expect: “What Is the Pomodoro Technique?”, “How the Method Works,” and “Benefits in Daily Work.” Claude kept the first two similar but turned the third into “Common Mistakes That Kill the Results”, a section that helps the reader instead of restating the obvious. The tools diverged on tone, too: ChatGPT hedged, suggesting you adapt the timer to 30- or 45-minute sessions, while Claude drew a hard line, “a pomodoro is either complete or it isn’t.” Same format, more opinion.

When precise instruction-following matters, both outputs still need checking against the brief. In this round both met the requested format, and ChatGPT’s neutral approach was a reasonable fit. I gave Claude a slight editorial edge for the usefulness of its final section, rather than for following more instructions.

Pricing: ChatGPT vs Claude

The June writing samples stay as they were. The prices below are a separate snapshot checked on October 21, 2026, in US dollars before applicable tax. Model access and usage limits can change without changing a plan’s name.

Individual plans in USD before tax, checked October 21, 2026. Annual Claude Pro billing is paid up front. Made with Meiku Charts.

Check the current ChatGPT Plus terms, ChatGPT Pro tiers and Claude plans before subscribing. Plus is USD 20 per month. Claude Pro is USD 20 monthly or USD 200 annually. Higher-usage options include ChatGPT Pro at USD 100, 200 and 500 per month, and Claude Max at USD 100 or 200.

Both tools offer a free plan. Claude’s current pricing page lists memory across conversations. OpenAI says ads may appear for eligible Free and Go users in selected regions, while Plus and Pro remain ad-free. Input limits, file handling and model availability vary by feature and account. See OpenAI’s current advertising guidance for eligibility.

A comparison to keep beside the samples

The writing observations belong to the June test. Current billing terms are shown separately from the historical model labels. Made with Meiku Charts.

Which should you use?

For these tasks, the choice depends on the draft you want to edit. The June results are useful examples to compare with your own prompts, rather than a ranking of every model released since then.

Choose Claude if you…

Want a starting draft with the sort of concrete details and varied rhythm Claude used in this test
Write blog posts, newsletters, or anything where voice and point of view matter
Value the free plan’s listed memory feature, subject to its current availability and usage limits

Choose ChatGPT if you…

Prefer the more restrained marketing sample shown here, while still checking every factual claim
Write to rigid formats or feed it data you can’t afford to have embellished
Want tools such as image generation alongside writing, rather than choosing on prose style alone

Both are general assistants. A purpose-built writing tool may suit a team that needs repeatable marketing workflows or SEO features. Compare the workflow and cost rather than assuming the specialist produces better prose. Our Jasper pricing breakdown and Writesonic review cover those alternatives.

Comfort with an app matters too. The more time I spend in one interface, the less natural the other feels. That preference can influence the choice without saying much about the quality of either tool’s writing.

Frequently asked questions

Is Claude better than ChatGPT for writing?

I preferred Claude’s tone in these June samples, while ChatGPT’s ad copy avoided the unsupported numerical claim. Both met the structured brief. That small comparison does not establish which current model is best for every writing job.

Is Claude free to use?

Yes. Claude has a free plan, and its current pricing page lists memory across conversations. Usage limits and available models can change. Sonnet 4.6 is the model recorded in the June comparison, not a permanent promise about the free plan.

Does ChatGPT show ads now?

OpenAI says ads may appear for eligible Free and Go users in selected regions. Its US test began on February 9, 2026. Plus, Pro, Business, Enterprise and Edu accounts are excluded. Availability is a rollout, rather than a claim that every free user sees ads.

Which is cheaper, ChatGPT or Claude?

Both entry paid plans cost USD 20 when billed monthly: ChatGPT Plus and Claude Pro. Claude Pro also offers USD 200 billed annually, about USD 16.67 per month over the year. The advertised USD 17 figure is rounded. Higher-usage plans are listed separately and are not necessary for every writing task.

Can ChatGPT and Claude write long articles?

Both can draft long articles, but the supported input length, file handling and usage allowance depend on the model and plan available to your account. A draft still needs structural editing and fact-checking. A larger allowance does not guarantee a more accurate article.

Which AI writes more like a human?

In the June samples shown here, I preferred Claude’s concrete details and uneven sentence rhythm. ChatGPT’s prose was smoother and more generic. The examples let you judge that preference, but neither writing style proves that a text was written by a human.

Originally published on Meikuio — ChatGPT vs Claude: Which Is Better for Writing?.

AI Agent Examples: What Is in Production and What the Numbers Mean

Egor Sivchik — Sun, 18 Oct 2026 19:50:57 +0000

By Ryan Kowalski. Originally published on Meikuio on July 18, 2026. Updated for this edition on October 18, 2026.

AI agent examples include Klarna’s customer-service assistant, coding tools such as Devin, and John Deere’s weed-targeting machinery. They do different jobs and have different levels of autonomy. The useful question is what a named system does in a real workflow, what evidence supports the result, and where a person still makes the decision.

Dated company figures below are historical claims, not independently audited measurements of current performance.

What is an AI agent, in one example?

An AI agent works toward a goal across several steps, using tools or taking actions as part of the process. A refund workflow might read an order, check a policy, propose or issue a refund within its permissions, and update the record. The label does not tell you how much of that workflow runs without review.

That distinction matters because vendors apply the word agent to different systems. Some are chat interfaces over existing automation. Others choose tools and actions dynamically. A deployed product can be useful without being fully autonomous. Check its permissions, stopping conditions and escalation path before comparing it with another example.

AI agent examples that are actually running in production

The examples below have named products or operators. That makes them more concrete than an anonymous demo, but it does not make their success rates comparable. Company reports, vendor case studies and independent experiments are different kinds of evidence. Keep the date, task and denominator attached to each number.

Company reports describe particular deployments, not comparable success rates. Table made with Meiku Charts, reviewed October 18, 2026.

These systems span assistants, coding agents and task-specific physical AI. Table made with Meiku Charts.

The clearest examples have bounded jobs. A support system can answer routine questions while escalating exceptions. A coding agent can work on an issue while a developer reviews the patch. Success in either case does not establish that the same product can run an entire department. Narrow scope makes the result easier to inspect.

Resolution rate needs a definition. It can refer to the subset of conversations offered to the agent, exclude certain requests, or use containment as a proxy for a successful answer. Ask whether the customer’s problem was solved, whether the customer came back, and how escalation was counted. A vendor-wide average is not a forecast for your own help desk.

Coding has unusually useful independent evidence. METR’s July 2025 randomized study found that experienced contributors took 19% longer with the early-2025 AI tools tested on their own repositories, despite expecting a speedup. The study involved 16 developers and 246 tasks. That is evidence about a specific setting and period, not a verdict on every coding agent or on October 2026 tools. METR’s February 2026 follow-up describes why selection effects and changing usage make a new estimate difficult.

The examples vendors show you (and what they leave out)

Vendor example pages often describe capabilities such as invoice processing or ticket resolution without identifying an operator, deployment date or measured result. Such a page can explain what a product is designed to do. It does not, by itself, show that a customer is using the entire workflow successfully.

Before treating a scenario as a deployed system, distinguish what has been demonstrated from what has been proposed. These common examples need that check:

A product-development agent that researches trends, checks rules and proposes formulations may be a concept or prototype. Look for a named operator and a record of the decisions it can actually execute.
A fully autonomous farm is a much broader claim than a machine that performs one field task. Verify which operations are deployed, which are supervised and which remain a proposed system.
A student or research prototype can demonstrate a mechanism without establishing a supported commercial product. Its evaluation conditions matter more than the label on a list of examples.

Concepts and prototypes can be useful, but they should be identified as such. A public demonstration, a paid pilot and a production deployment answer different questions. Ask who operates the system, since when, on which tasks, and with what oversight. Missing evidence is a reason to withhold a conclusion, not proof that the idea cannot work.

Klarna: workload, savings and human support

Klarna’s published workload figure makes it a useful case for separating volume, savings and service quality.

In February 2024, Klarna reported that its assistant had handled 2.3 million conversations in its first month, about two-thirds of its customer-service chats. The announcement equated that workload with 700 full-time agents and projected a $40 million profit improvement for 2024. Those were company claims. The profit figure was a projection, and workload equivalence was not a count of people dismissed.

Bloomberg reported on May 8, 2025 that Klarna planned more human customer-service options after emphasizing AI-led cost cuts.

The customer-service question does not end with a workload figure. Klarna’s published support and complaints routes still include human contact. A useful assessment has to measure quality, repeat contacts and escalation as well as cost. Keeping people available can be part of a working deployment, rather than evidence that the underlying automation was never useful.

The same test applies to an AI app-builder or outsourced development service. Ask which work the software performs, which work people perform, and how the handoff is disclosed. A polished interface does not establish autonomous execution. Buyers need a reviewable record of the work and a clear account of who is responsible when it goes wrong.

How many of these actually work?

Case studies establish that particular workflows exist. They do not establish the success rate of all agent projects, because unsuccessful or unfinished projects are less likely to become public examples.

Gartner, in June 2025, projected that more than 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear value, and weak risk controls. Gartner also reckoned that of the thousands of vendors claiming to sell “agents,” only around 130 were doing something real, and called the rest “agent washing.”

McKinsey’s November 2025 survey separates adoption stages: 23% of respondents said their organizations were scaling an agentic AI system somewhere, and another 39% said they were experimenting. Together that is 62% reporting some agent use. The survey’s roughly 6% AI high-performer group met a broader definition involving significant value and at least 5% EBIT impact from AI. It is not the share of companies making any profit from agents, and it is not a conversion rate from the 23% group.

Share of survey respondents, November 2025. Any use includes experimentation and scaling, so the bars overlap. Chart made with Meiku Charts.

The two bars overlap: organizations scaling agents are included in the 62% reporting any use. They are not slices of a pie. The separate high-performer measure covers AI more broadly and is deliberately not plotted as an agent-profit result.

For coding, benchmark scores and workplace productivity measure different things. A repository task in a study includes context, review and integration that a headline benchmark may not capture. Measure completion time, correctness, rework and review effort on your own tasks. A rising score does not remove those costs from a production workflow.

The working examples remain useful. Their limits are part of the example: what was attempted, what was counted, and how much supervision the result required. A list that includes only successful vendor case studies cannot tell you how often a new project will succeed.

AI agent examples you can actually run yourself

A small business can start with a bounded workflow rather than copying an enterprise deployment. The practical examples below are jobs to test, not a promise that installing an agent will make them reliable automatically.

It reads your incoming mail, sorts it, and drafts replies you approve before they send (email triage).
It sits on a call, transcribes it, and pushes action items straight to your task list (meeting notes).
It scores inbound leads against your criteria and fills in the missing contact details (lead qualification).
It answers routine customer questions from your help docs before a human sees them (support deflection).
It pulls line items out of invoice PDFs and files them as structured records (data extraction).

You can use a purpose-built application or connect tools in a workflow builder. A ready-made product can reduce setup time, while a custom workflow gives you more control over the steps. Compare usage fees, maintenance and review time. A visual builder is not automatically cheaper once those costs are included.

Set approval rules before connecting an agent to customers or payments. For example, it can draft a reply or propose a refund, then wait for a person to approve the action. Record failures and escalations alongside completed tasks. Start with one job whose output you can check, then expand only when the results support it.

Frequently asked questions

What is an example of an AI agent?

A support agent is a clear example when it can read a request, consult account information and take a permitted action. A system that only answers from help documents is a narrower assistant. Klarna and Intercom provide deployed support examples, while coding agents and task-specific physical AI illustrate other forms of action.

What is the most successful real AI agent example?

There is no defensible single winner across support, coding and physical machinery. They do different jobs, and most public success rates are company claims with different denominators. Choose a measure for the task: verified resolution, accepted code, reduced rework or an observed field outcome.

Are AI agents actually being used in 2026, or is it hype?

Deployed products exist, including the support and coding systems described here. Their existence does not establish the success rate of the wider market. Gartner’s 2025 cancellation forecast concerns projects through the end of 2027, not an observed count of failures today. A pilot should be evaluated separately from a production rollout.

What’s the difference between an AI agent and a chatbot?

A conversational interface can answer a question without taking an action. An agentic workflow uses tools or other steps to pursue a goal within its permissions. The two can share the same chat interface, so inspect the workflow rather than relying on the product’s name. Human approval can still be part of an agent’s design.

Can a small business actually use an AI agent?

Yes, for bounded jobs such as email triage, meeting notes, lead qualification, support and invoice extraction. Start with a task whose outputs you can verify. Include approval for consequential actions, track errors and review effort, and compare the result with the process you already use.

Do AI agents really save money?

They can, but savings need to be measured after software fees, setup, upkeep, errors and human review. A vendor’s resolved-conversation rate is not a profit calculation. McKinsey’s 6% high-performer figure is a broader AI category, not a measured percentage of companies earning any return from agents.

My Take

The examples are most useful when they show a bounded job and leave enough evidence to examine the result. A support reply, a code patch or an invoice record can be checked. A claim that an agent will run an entire operation needs much more evidence than a successful demonstration.

I would choose one repetitive job, record how long it takes today, and test the agent against that baseline. Count the time spent checking and fixing its work. Keep a person responsible for actions involving customers or money. Expand when the record supports it, not when a vendor’s example happens to look like your business.

Sources and measurement notes

Klarna’s 2024 announcement · McKinsey’s 2025 survey · METR’s 2025 controlled study · METR’s 2026 follow-up · Bloomberg’s May 2025 report · Salesforce’s Help-site case study · Sierra’s WeightWatchers case study

Other company records reviewed include Morgan Stanley’s June 2024 Debrief announcement, Commonwealth Bank’s 2026 half-year presentation, current coding-product documentation, and John Deere’s See & Spray documentation. Product capabilities and historical reported results are not independent audits.

The original Meikuio article contains the interactive version.

Have I Been Pwned Alternatives: What’s Free in 2026

Egor Sivchik — Tue, 06 Oct 2026 08:14:39 +0000

By Marcus Hale. Originally published on Meikuio on August 7, 2026. Reviewed for syndication October 6, 2026.

The catalogue comparison and research observations below were recorded on August 2, 2026. Prices and product documentation were rechecked for this edition; historical counts are labelled as snapshots.

The best Have I Been Pwned alternative depends on what you are actually trying to replace. If you want to check your own email address, you do not need one, because that check costs nothing. If you need a phone number, a name, or an API, three specific tools do things Have I Been Pwned does not. Here is what each one runs on.

I spent a day reading the documentation behind the tools on the current lists, instead of their marketing. Prices came off the vendors’ own pricing pages. Provenance came out of their own documentation, and where a company does not say where its data comes from, I wrote that down too. Two of the tools that search engines still recommend are dead. One of the most-recommended alternatives is the same database with a different logo, and I can show you that rather than assert it.

What are you actually looking for when you search for a Have I Been Pwned alternative?

Three different people type this phrase, and they need three different answers. The first wants to check an email address and believes the check costs money. The second wants something the service does not do at all, like searching a phone number or a Social Security number. The third is a developer or an administrator who hit the API price and wants a cheaper way to query at volume. Only the second and third have a real problem.

That first group is looking for an alternative to something that is already free for them. That is not their fault. The paid plans are advertised on the same site as the free search, the pricing page is long, and the top of it runs to five figures a year. If you glance at that page you would reasonably conclude the whole thing is a subscription product.

It is not. A basic email lookup needs no login, and the vendor lists what costs nothing in plain language on its own subscription page. I will quote that list in a moment, because it settles the misconception this whole search is built on.

Most people searching for an alternative are trying to replace something that is already free for them.

The second group has a gap nothing else fills. Have I Been Pwned indexes email addresses, and after the leaks of the last two years a lot of people want to search a name or a phone number instead. One free tool does that, and it holds a completely different set of records. I come to it below, with the count printed on its own front page.

The third group is being priced by design. Read the plan descriptions and the intended customer is named outright: organisations, managed service providers, enterprises. Nobody at the company is pretending the paid tiers are for individuals.

What does Have I Been Pwned cost, and who is it charging?

Checking your own email address costs nothing. The subscription page prints its free tier as a list, and that list reads: browser email search, email notifications, Pwned Passwords, breach monitoring, and monitoring domains with up to 10 breached addresses. Everything a person doing a personal check needs is inside that sentence. I read the page on 2 August 2026.

What you pay for is volume and machinery. The paid ladder is priced by requests per minute, by how many breached addresses a domain can hold, and by how many domains you watch. Here is what the vendor prints, in its own words and figures.

Have I Been Pwned Commercial Pricing: Tier Limits, Rates, and Domain Capacities

Data snapshot: Have I Been Pwned official enterprise pricing portal (August 2026)

Free: Price as printed: $0; Terms: Requests per minute: n/a; Terms: Domains: up to 10 breached addresses; Who the vendor says it is for: Anyone checking an email address
Core 1: Price as printed: $4.39/mo billed as $52.68/y; Requests per minute: 10; Domains: 1; Who the vendor says it is for: "organisations monitoring a small number of their own domains"
Core 5: Price as printed: $319/mo billed as $3,828/y; Requests per minute: 1000; Domains: 20; Who the vendor says it is for: "organisations monitoring a small number of their own domains"
Pro 1: Price as printed: $379/mo billed as $4,548/y; Requests per minute: 1000; Domains: 50; Who the vendor says it is for: "MSPs and larger enterprises"
Pro 5: Price as printed: $4,599/mo billed as $55,188/y; Requests per minute: 16000; Domains: 800; Who the vendor says it is for: "MSPs and larger enterprises"
High RPM 24000: Price as printed: $5,833/mo billed as $69,996/y; Requests per minute: 24000; Terms: Domains: none; Who the vendor says it is for: "high-throughput API access"

Read the right-hand column again. Organisations, managed service providers, enterprises, throughput. Not one of the three paid families is described as a consumer product, and the description is the vendor’s, not mine. The pricing page is a business page that happens to sit next to a free tool.

Two features matter if you are the third kind of searcher. HIBP now lists k-anonymity email search on Pro and High RPM, while Core uses direct email lookup. Pro supports domain stealer-log access; Core does not. Check the current plan comparison against the exact endpoint you need rather than assuming every paid tier includes the same data.

Plan names changed recently, which is why some guides quote prices that no longer exist. The support page says it plainly: in late March 2026 the original Pwned 1 to 5 and Pwned Ultra plans were replaced by Core, Pro and High RPM, and old plans are no longer available for new purchases. Existing subscriptions, the same page says, “will continue to function as usual until August 2, 2026, when they’ll rollover to a corresponding new plan at the next renewal after that date.”

For scale, the entry tier was 10 requests per minute for $3.50 a month when this pricing model was introduced in November 2022. It is $4.39 now. The top of the ladder is close to $70,000 a year.

Where does each tool’s breach data actually come from?

Every list I measured skips this question, and it decides whether a tool is an alternative at all. A breach checker is a search box on top of a corpus. That is all it is. If two search boxes query the same corpus, switching between them changes the interface and nothing else. So I went through the documentation of every tool on the current lists and sorted them by where the records come from. None of the lists does this. It takes an afternoon.

Breach Checker Provenance: Which Tools Mirror HIBP vs. Query Independent Databases

Data snapshot: Vendor platform documentation and API disclosures (August 2026)

Mozilla Monitor: Whose records it searches: Have I Been Pwned; Free without an account?: Search yes, alerts need an account; Stated by: Mozilla's own FAQ
1Password Watchtower: Whose records it searches: Have I Been Pwned; Free without an account?: Paid subscription; Stated by: 1Password support docs
Bitwarden breach report: Whose records it searches: Have I Been Pwned; Free without an account?: Yes, all plans; Stated by: Bitwarden help pages
Apple Passwords: Whose records it searches: Apple's own set; 1.5 billion was a historical documentation figure, not a 2026 measurement; Free without an account?: Built in; Stated by: Apple platform security guide
Chrome Password Checkup: Whose records it searches: Google's own set, over 4 billion credentials; Free without an account?: Built in; Stated by: Google security blog
DataBreach.com: Whose records it searches: Its own, 27.3 billion records; Free without an account?: Yes; Stated by: Printed on its own front page
Cybernews leak checker: Whose records it searches: Its own, more than 15 billion accounts; Free without an account?: Yes; Stated by: Its own FAQ
XposedOrNot: Whose records it searches: Its own, open source; Free without an account?: Yes; Stated by: Its own FAQ and API docs
LeakCheck: Whose records it searches: Its own, more than 10 billion records; Free without an account?: Sources only, records are paid; Stated by: Its own wiki
DeHashed: Whose records it searches: Its own; Free without an account?: Search is free, results are paid; Stated by: Its own site

The pattern in that table is not the one I expected before I started reading. The two checkers that ship inside an iPhone and inside Chrome, already installed on devices most people are holding, do not use Have I Been Pwned at all. Apple compares your saved passwords against its own curated set, described in Apple’s security documentation as approximately 1.5 billion passwords. Google warns you when a saved login appears among what its own security blog calls over 4 billion credentials that Google knows to be unsafe. Two separate corpora, neither of them the one everybody argues about.

1Password Watchtower uses HIBP’s Pwned Passwords service with a hash-prefix lookup. Bitwarden uses HIBP for its data-breach report, which is available on all plans. These are distinct checks: searching an email address in a breach and checking a saved password against compromised-password data do not answer the same question.

So the honest summary of the category is short. Free browser and phone checkers run their own data. Paid password managers rent someone else’s. And one very popular free website is a front end for the exact database it is listed as an alternative to.

Which alternatives are Have I Been Pwned with a different logo?

Mozilla Monitor is the first name on almost every list of alternatives, and Mozilla does not hide what it is. Its support FAQ answers the question in one sentence: “Mozilla Monitor gets its data breach information from a publicly searchable source, Have I Been Pwned.” The same page says elsewhere that a breach may be missing from Monitor because “Have I Been Pwned, our breach source, hasn’t been granted access” to it. That disqualifies the tool as an alternative in any meaningful sense. It is a different front door to the same building.

I wanted to see that rather than take anyone’s word for it, so I opened both breach databases on the same afternoon and counted. Have I Been Pwned prints its own totals on its breach list page: 1,022 total breaches and 17.8 billion pwned addresses. Mozilla Monitor’s breach database page held 1,023 individual records. The five most recent entries were identical on both sites, in the same order, with the same dates added: SplitVPN on 1 August 2026, Houston City College on 28 July, Suno and Paidwork on 20 July, Fluke on 15 July.

Two catalogues, the same five entries

Data snapshot: Both breach databases read 2 August 2026

Have I Been Pwned: Total breach records: 1022; Counting basis: Printed total; 17.8B pwned addresses; Five newest entries: SplitVPN — 1 Aug; Houston City College — 28 Jul; Suno — 20 Jul; Paidwork — 20 Jul; Fluke — 15 Jul 2026
Mozilla Monitor: Total breach records: 1023; Counting basis: Count of breach-record links; HIBP is its stated data source; Five newest entries: SplitVPN — 1 Aug; Houston City College — 28 Jul; Suno — 20 Jul; Paidwork — 20 Jul; Fluke — 15 Jul 2026

The one-record gap comes from two counting methods, not evidence of independent data. All five names, order and addition dates matched.

One record apart, and that gap is inside the margin between two ways of counting, so I am not going to build anything on it. The point is the five names. Same breaches, same order, same day they were added.

None of this makes Mozilla Monitor bad. It is free, it sends alerts, and after the paid tier closed Mozilla expanded free monitoring to cover up to 20 email addresses per user, which is more addresses than most people have. Use it if you like the interface. Just do not switch to it believing you are checking a second database, because you are not.

Same five newest breaches, same order, same dates added, on both sites, on the same afternoon.

Look further down the lists and it repeats. One of the highest ranking guides for this exact search puts its own product at number one and then describes it, in the same article, as a tool that “uses verified breach intelligence (including Have I Been Pwned)”. The same article calls Mozilla’s product a “free breach checker built on top of Have I Been Pwned”. Two of its five recommendations are the thing it is recommending alternatives to, and the article says so itself.

Which free breach checkers are actually worth using?

Three tools hold their own records, cost nothing, and answer a question Have I Been Pwned does not. Those are the ones that earn the phrase Have I Been Pwned alternative, and they come first below. Two more appear on the same lists and are free only in a narrow sense, so they get their own heading at the end rather than a recommendation. I have noted what each one asks for before it tells you anything.

DataBreach.com, if you need to search a name or a phone number

This is the one that fills the gap. Have I Been Pwned indexes email addresses. DataBreach.com, built by Atlas Privacy, takes an email, a name or a phone number, and its front page prints the size of what it is searching: 27,367,392,050 leaked records. The page also carries the line “Anonymous search, we do not store any search data” directly under the box.

Its records are different, and that is checkable in about a minute. On the afternoon I looked, its five newest entries were Novum Energy at 227,000 rows, Campbell University at 1.1 million, Quest Health Solutions at 2.1 million, Live Casino at 1.8 million, and Bank of Baroda at 39 million. Those were added on 31, 30, 29, 28 and 27 July 2026. Not one of them appears among the newest entries on Have I Been Pwned. Two corpora, no overlap at the recent end.

The original research snapshot on 2 August 2026 recorded 27.3 billion records. On 6 October 2026, DataBreach.com advertised 27,810,011,995 records. These are vendor-reported record counts, not unique people or independently measured coverage. Check the current front page rather than recycling an older headline.

XposedOrNot, if you want an open source second opinion

XposedOrNot holds its own aggregated corpus and publishes its code. Email and password checks are free, no account is needed for an email lookup, and the project states that its API “will continue to remain completely free of charge” with the application and related files open source on GitHub. Domain queries are the exception and need an API key issued only to a verified domain owner.

It is the closest thing in this category to an independent second opinion, because you can read how it works rather than trust a claim about it. The interface is plainer than the commercial tools. That is the trade.

Cybernews leak checker, if you want to check a phone number quickly

Run by a security publication rather than a security vendor, and it takes an email or a phone number with no account required. Its own FAQ describes a database of more than 15 billion compromised accounts, containing “nothing more than the email address and the source of the leak”, stored hashed with bcrypt. That narrowness is a feature. There is very little for anyone to lose if the checker itself is ever breached.

Before you type your address into anything

Every tool on this page asks you to hand over the identifier you are worried about. That is unavoidable for a lookup, but it is not free of cost, so notice which ones make you create an account first. The ones that do not, on my reading, are DataBreach.com, Cybernews, XposedOrNot, and the basic email search on Have I Been Pwned itself. Mozilla sits in between: its own FAQ says you may search an address without signing up, and the account is what you need for ongoing alerts. Several commercial checkers require an account and a payment method before showing you anything at all.

DeHashed and LeakCheck, and why “free” needs a footnote

Both appear on consumer lists as free tools. Both are free in a specific and limited way. LeakCheck’s public API returns “only the list of breach sources and the categories of exposed data, never the data itself”, and the actual records need a paid plan. DeHashed lets you run a search without paying, then requires a subscription to view what the search found. That is a demo, not a free tier.

Neither is built for you. DeHashed searches by IP address, phone, VIN and more, and its own description names security analysts, journalists and investigators. They are good tools aimed at somebody else.

Have you tried the breach check you already own?

There is a breach checker inside your phone, your browser and your password manager. They run automatically, they check the passwords you actually use rather than an address you type from memory, and you have already paid for them or they came free with the device. Not one of the lists I read mentions them. They should.

On an iPhone or a Mac, saved passwords are compared against Apple’s own set of leaked passwords using what Apple’s security documentation calls a form of cryptographic private set intersection, with only a 15 bit prefix of a hash leaving the device. In Chrome, Password Checkup warns you when a saved login is among the credentials Google knows to be unsafe, using a mix of anonymised lookup and private set intersection. In both cases the comparison is done in a way that does not hand your password to anyone, and in both cases the corpus belongs to the platform. Neither asks Have I Been Pwned anything.

1Password Watchtower checks saved passwords against HIBP’s Pwned Passwords data. Bitwarden offers an email data-breach report on all plans, including Free. Its exposed, reused and weak-password reports are separate vault-health features, available to Premium users and paid organisations. Do not confuse a free email breach lookup with a full saved-password audit. Bitwarden documents the distinction here.

The advantage here is not the size of the database. It is that these tools know which passwords you use and where, so the answer arrives as a list of accounts to fix rather than a list of company names you half remember signing up to. That is the difference between being informed and being able to act. Apple and Google provide these checks at no extra charge. If you use a separate password manager, check which reports your plan includes.

What about the free monitoring attached to your credit card?

Several US card issuers include dark web and identity monitoring at no cost, and two of them do not require you to be a customer at all. That covers different ground from a breach checker, because it watches a Social Security number and a credit file rather than a list of hacked websites. I have written about which issuers give what, and how the free tools compare on score type and bureau, in the piece on best credit monitoring service, and about whether any of it justifies a paid subscription in is identity theft protection worth it. If you have a card in your wallet, start there before you buy anything.

One question about it goes unanswered everywhere, and it is the same question this article asks of everything else: whose records are those issuer alerts actually searching? A bank does not build a breach corpus. It licenses one. None of the issuers I looked at names the supplier, and that silence tells you how much they expect you to ask.

Does dark web monitoring work, or is it theatre?

It tells you something has already happened. That is the whole of it. A monitoring service cannot remove your data from anywhere, cannot stop a record being resold, and cannot undo a leak. It can shorten the gap between the leak and your knowing about it, which is worth something, and it is a smaller something than the advertising implies.

A 2019 Consumer Federation of America survey found that 36 percent of respondents exposed to dark web monitoring ads thought the services could remove information, while 37 percent thought they could prevent purchased data being used. The organisation warned that alerts cannot reverse an exposure. The date matters: these figures describe that survey, not a new 2026 poll.

The advertising deserves scrutiny. In the original research I did not locate a category-wide US regulatory guide specifically evaluating dark web monitoring claims. That search does not establish that none exists. The FTC warns that unsolicited emails claiming your data is for sale may themselves be phishing, while the Consumer Federation of America directly examined consumer misunderstanding of these services.

Whether the paid version of this is worth buying is a longer argument, and I made it separately in is identity theft protection worth it. The short version for this page: the alert is the product, and the alert arrives after the fact.

Which tools on these lists are not sold to you?

Open almost any list of Have I Been Pwned alternatives and you will find Flare, Breachsense, Intelligence X or SpyCloud in it. None of those companies is selling to you. They sell to security teams, and their pricing, their onboarding and their dashboards all assume there is a team on the other end.

You can tell from how they describe themselves. Breachsense positions itself as built “for the security-team use case that HIBP was never designed for”, monitoring continuously and integrating with a stack. Flare talks about tracking exposure across clear web and underground sources rather than a point-in-time check. Intelligence X describes itself as a search engine and data archive covering Tor, I2P and data leaks. Those are procurement pitches.

There is nothing wrong with any of them. They are simply in the wrong list. A person who typed a worried question into a search box does not need a threat exposure management platform, and putting one in front of them is how a list gets to twelve entries when the honest number is four.

The other way lists get padded is by counting annual subscriptions as a Have I Been Pwned alternative. One roundup of twelve puts nine paid identity suites on it, which is a different product answering a different question. If that is what you are shopping for, read the individual reviews instead of a list that treats a free lookup and an annual subscription as substitutes. I have taken two of them apart at length, the Identity Guard review and LifeLock’s dark web monitoring in particular.

The same padding shows up elsewhere. One page ranking on the first screen for this search lists its top alternatives to a breach checker as an email deliverability tool, an anti-detect browser and a file sync utility. None of the three checks a breach. Those recommendations do not answer the breach-checking question, whatever process produced the list.

Which recommendations in the search results are out of date?

Two products that these lists still recommend no longer exist. Google’s dark web report is gone. Mozilla’s paid Monitor Plus tier is gone. Both had closed by the August research snapshot, and some articles I checked then had not caught up.

Google’s own support page gives the dates without ambiguity: “January 15, 2026: The scans for new dark web breaches stop” and “February 16, 2026: The dark web report is no longer available.” On that second date, Google says, all data related to the report was deleted. The reason it gives is unusually frank for a shutdown notice: “While the report offered general information, feedback showed that it didn’t provide helpful next steps.”

Now look at what is still on the internet telling you to use it. Gen Digital, the company that owns Norton and LifeLock, publishes a consumer guide to finding out whether your information is on the dark web. It carries a section headed “Google’s free dark web checker” and tells the reader: “You can set up dark web monitoring via your Google account for free on the Google Dark Web Report page. To qualify for dark web scanning, you must have a regular consumer Google account.” That page is stamped Published: January 07, 2026, which is three weeks after Google announced the closure and eight days before the scanning stopped. It still said it when I read it on 2 August 2026.

Note

To be precise about what that is and is not: this is Norton’s own guide, not a page ranking in the top ten for the searches this article targets. I checked. The point is not where it ranks, it is that a security company with a research budget has been recommending a deleted product for seven months.

The smaller version of the same problem is a name. Mozilla renamed its product from Firefox Monitor to Mozilla Monitor, and articles written after the rename still use the old one. A widely read roundup dated 29 December 2025 gives “Firefox Monitor” a section, a feature list and a pricing block, and never uses the current name once. A guide dated 28 January 2026 does the same. A large crowd-sourced alternatives page, last touched on 31 October 2025 and still ranking, contains the words discontinued, shut down and no longer exactly zero times, while listing entries whose own dates go back to 2018.

Every Have I Been Pwned alternative list I measured carried at least one of these faults: a dead product, a retired product name, a page that never mentions whether anything on it still runs, or entries from a different software category altogether. Check the date on any list you read. Then check whether the thing it recommends still answers when you knock.

What can these tools miss?

Two important sources of stolen credentials are company breach dumps and infostealer logs from infected devices. Free breach checkers can differ in how much of each they expose. A leaked company database and an infected endpoint need different responses.

Have I Been Pwned describes the second kind precisely: stealer logs “are the result of malicious software running on infected machines that collect email addresses, passwords and the website they’re entered into at login”. That last clause is the difference that matters. A breach dump tells you a company lost your address. A stealer log tells someone your password and the exact page you typed it into, plus, often, a live session cookie that skips the password entirely.

The scale is not small. When one stealer log corpus was processed into Have I Been Pwned it contained 183 million unique email addresses, of which 91 percent were already known and 16.4 million had never been seen before. That leaves 9 percent of an enormous load as material nobody had on file. A separate credential stuffing corpus loaded soon after held 1,957,476,021 unique email addresses. It also held 1.3 billion unique passwords, and 625 million of those passwords were new to the service. Those are the operator’s own published figures.

Here is the catch for anyone shopping on price. Core does not include HIBP stealer-log data. Pro includes access for verified domains. Individuals can check exposure involving their own verified address. The paid domain and API capabilities are separate from the free personal lookup; consult HIBP’s current endpoint and plan documentation before choosing a subscription.

One limitation applies to this article too. I did not find a vendor-neutral comparison establishing which of these corpora is most complete. Different record counts and different recent entries do not rank coverage. The August comparison below establishes provenance and overlap at that snapshot, not a universally best database.

Why do you keep seeing “billions of passwords leaked”?

Large credential headlines often describe compilations of earlier leaks, sometimes with duplicates, rather than a newly compromised company. Read the researchers’ description of each dataset before interpreting its record count.

Take the 16 billion credentials story from June 2025. The original research described records “scattered across 30 different databases” and noted that the raw count “includes duplicates, as is common in these types of compilations”. Other outlets corrected the framing quickly, one stating flatly that “this is not a new data breach, or a breach at all”. RockYou2024 was a file of just under 10 billion plaintext passwords assembled from earlier leaks. The 26 billion record collection from January 2024 was described by its own discoverers as records from thousands of compiled and reindexed leaks.

This matters when you are choosing a tool, because a checker that ingests every compilation will show you more hits without telling you anything more. The stealer log load described earlier shows the arithmetic: nine records in ten were already known, and the tenth was the only one that told anyone anything.

Compromised-password datasets also support password security standards. NIST SP 800-63B-4 calls for checking newly chosen passwords against a blocklist of common, expected or compromised values. Previously breached passwords are one example. It does not require a particular vendor or explain why HIBP chose to make Pwned Passwords free. The free service is useful infrastructure; the standard and the vendor business model are separate facts.

Who keeps Have I Been Pwned running?

One person, essentially, and that is the strongest argument for having a second tool bookmarked. The service was put up for sale in 2019. The sale collapsed, and the announcement in March 2020 was blunt: “Have I Been Pwned is no longer being sold and I will continue running it independently.” I could not find a published succession plan or continuity arrangement anywhere. That absence is not an accusation, it is just a fact you should hold when you decide how much of your security routine to hang on one site.

There is a detail in the breach list that says more about how the thing is run than any policy page would. In March 2025 a phishing attack got into the operator’s own Mailchimp account and exported his blog newsletter list. The breach page reads: “The exported list contained 16k email addresses and other data automatically collected by Mailchimp including IP address and a derived latitude, longitude and time zone.” It is filed under 16.6 thousand affected addresses, and it was added to the database on 25 March 2025.

He put his own breach in his own catalogue inside a month, with the same fields and the same treatment as everyone else’s. That is not nothing. It is also not a succession plan. Keep a second tool bookmarked.

How do you check whether your data has leaked?

Five steps. The email lookups and built-in Apple and Google checks are free; paid password-manager features are optional. Doing them out of order is how people end up paying for something they already had.

Search your email address on Have I Been Pwned. No account, no payment. This is the check most people are looking for an alternative to, and it costs nothing.
Search the same address, plus your name and phone number, on DataBreach.com. Different corpus, different fields, and this is where you find things the first search structurally cannot see.
Open the password checker you already own. Passwords on an iPhone or Mac, Password Checkup in Chrome, Watchtower in 1Password, or the exposed-password report on an eligible Bitwarden paid plan. Bitwarden’s free data-breach report searches your email address; it does not audit saved passwords. A password-level check helps identify accounts to fix.
Turn on alerts, once. Notification on Have I Been Pwned or a Mozilla Monitor account will email you when a new breach lands. Both draw on the same data, so pick one, not both.
Fix by severity, not by date. Anything that exposed a password you reused goes first. Anything that exposed only an address can wait, and mostly needs nothing but a sharper eye for phishing.

If a Social Security number is the thing worrying you, these email lookup tools are not a complete answer. A free credit freeze at each of the three major US bureaus helps prevent new-account fraud, but it does not stop every form of identity misuse.

So which Have I Been Pwned alternative should you use?

By situation, because there is no single winner and anyone crowning one is selling something.

Breach checkers: accepted identifiers

Data snapshot: Own search forms and documentation, read 2 August 2026

Have I Been Pwned: Email address: Free; Name: Not supported; Phone: Not supported; Password: Free; Checks one password entered by the user
DataBreach.com: Email address: Free; Name: Free; Phone: Free; Password: Not supported
Cybernews: Email address: Free; Name: Not supported; Phone: Free; Password: Not supported
XposedOrNot: Email address: Free; Name: Not supported; Phone: Not supported; Password: Free; Checks one password entered by the user
Mozilla Monitor: Email address: Free; Account required for alerts; Name: Not supported; Phone: Not supported; Password: Not supported
Apple Passwords / Chrome Checkup: Email address: Not a search field; Name: Not supported; Phone: Not supported; Password: Built in; Compares saved passwords on the device
DeHashed / LeakCheck: Email address: Paid results; Searchable; payment needed to view records; Name: Not established; Phone: Paid results; Searchable; payment needed to view records; Password: Not supported

Domain search is a business feature, excluded because it was not tested. A missing field can reflect what the database indexes, rather than a missing product feature.

You want to check an email address. Use Have I Been Pwned. You do not need an alternative, and it costs nothing.
You need to search a name or a phone number. DataBreach.com, free, no account, and a separate set of records.
You want an open source second opinion. XposedOrNot.
You want the check attached to your actual passwords. Apple Passwords or Google Password Checkup; Bitwarden’s exposed-password report requires an eligible paid plan. Its free breach report searches an email address instead.
You want alerts and nothing else. Mozilla Monitor, up to 20 addresses, with the understanding that it is the same database underneath.
You are a developer who hit the API price. Nothing here replaces it cheaply for domain monitoring at scale. XposedOrNot’s free API is the closest, and it is not the same product.
You are protecting a company. You are the customer the paid tiers were written for, and the enterprise platforms named above are built for exactly that.

If you are here because a company just emailed you about a breach, do not start with a checker at all. Secure the email account first, then change the password on the breached service and anywhere you reused it. The checkers tell you what else is exposed, which is the second job, not the first. For what one of these incidents looks like from the inside, including how quickly the records reached a public catalogue, I went through a recent one in the Aura review.

Frequently asked questions

Is Have I Been Pwned free?

Yes for the thing most people want. The subscription page lists browser email search, email notifications, Pwned Passwords, breach monitoring and monitoring domains with up to 10 breached addresses as costing nothing. Paid plans start at $4.39 a month billed as $52.68 a year, and the vendor describes them as products for organisations, managed service providers and high-volume API users.

Is there a better alternative to Have I Been Pwned?

For a basic email lookup, HIBP remains a useful free starting point. For name or phone searches, DataBreach.com offers fields HIBP does not; its advertised corpus is now over 27.8 billion records. Mozilla Monitor queries HIBP data, so it offers a different interface and alerts rather than an independent corpus. No record total alone establishes which checker has better coverage.

Does dark web monitoring work?

Monitoring can alert you that details have appeared in a dataset; it cannot undo the leak or prevent all subsequent misuse. In a 2019 Consumer Federation of America survey, 36 percent of respondents who had seen dark web monitoring ads mistakenly thought the services could remove information from the dark web. Treat that as a dated survey, not a measure of consumer beliefs in 2026.

How do I check if my SSN is on the dark web or is being used?

An email breach lookup is not an SSN misuse check. HIBP does not offer an SSN search, and a name or phone lookup is not equivalent. Some US card issuers provide identity monitoring, discussed in our credit monitoring guide. A free credit freeze at Equifax, Experian and TransUnion restricts access to your credit file and helps prevent new-account fraud. It does not stop every form of identity theft or misuse of existing accounts.

Was my data leaked, and how would I find out?

Search your address on Have I Been Pwned, then search your address, name and phone on DataBreach.com. Between them you cover two separate corpora. Then open the password checker in your browser or password manager, which compares the credentials you actually use rather than the ones you remember having.

What is the best free dark web scan?

Three of them cover the ground: Have I Been Pwned for email, DataBreach.com for other identifiers, and XposedOrNot for an independent second look. Watch the word free on commercial tools. DeHashed lets you search without paying and then requires a subscription to see results, and LeakCheck’s free tier returns the names of breaches but not the data inside them.

Can I check a phone number for breaches?

Yes, on two of the tools here. DataBreach.com takes a phone number alongside email and name, and the Cybernews leak checker accepts a phone number with no account. Have I Been Pwned does not, because it indexes email addresses and usernames.

Is Have I Been Pwned safe to use?

Its password checking is built so the service never receives your password. Pwned Passwords hashes locally and sends only the first five characters of the hash, and it is free and requires no login. The honest caveat has nothing to do with security: it is run independently by one person, the 2019 sale process ended with no buyer, and no succession plan has been published.

Why do some breaches not show up in a search?

Some breaches are marked sensitive and require verification of the email address before results are shown. Some datasets have been retired, and no checker includes every incident. The original August research counted 85 sensitive breaches and two retired ones; those counts are a historical snapshot, not current totals. A second independently collected database can add context, but a negative result is not proof of safety.

What should I do if my email shows up in a breach?

Start with your email account, because whoever controls it can reset everything else: change that password and switch on two factor authentication there. Then change the breached account’s password, then every other place you reused it. If the breach exposed only an address, expect more phishing and change nothing. If it exposed a password you used elsewhere, treat every one of those accounts as compromised until you have changed it.

My Take

I went looking for a Have I Been Pwned alternative and found that most of them are the same database in different clothes, and that the two checkers already sitting in Apple’s and Google’s password managers are the ones nobody counts as alternatives at all. The category is smaller than the lists suggest. Three free tools do separate work, and one of them exists because Have I Been Pwned deliberately does not index the fields people are most frightened about.

What I would do, and did: search the email address on Have I Been Pwned, search the name and phone on DataBreach.com, then open the password checker already sitting in the browser and fix what it flags. That takes ten minutes and costs nothing. I would not buy a subscription off the back of a breach alert, and I would not switch to a tool that queries the same records I just searched. If a Social Security number is the worry, stop reading lists and freeze your credit files instead.

Sources and original

Read the original Meikuio article and interactive comparisons.

How to Spot an AI Voice Scam (Before It Costs You)

Egor Sivchik — Mon, 05 Oct 2026 07:57:23 +0000

By Marcus Hale. Originally published on Meikuio on June 6, 2026; reviewed for syndication October 5, 2026.

If a call sounds exactly like your child, your boss, or your bank, that is no longer proof of anything. The fastest way to spot an AI voice scam is to stop trusting the voice and verify through a separate channel you control – hang up and call back on a number you already have. Some commercial tools advertise voice cloning from as little as three seconds of audio, while the FBI’s 2025 report documents substantial AI-related fraud losses. Here is how these calls work and the exact protocol that stops them.

The red flags, at a glance

Most AI voice scams share the same fingerprints. None of them is about how the voice sounds – the giveaway is what the caller wants and how fast they want it.

Manufactured urgency: “I need this in ten minutes.” Pause and verify.
Unusual payment method: Gift cards, cryptocurrency, wires, or a “confidential” account.
Secrecy: “Don’t tell Mom.” The caller wants to prevent a second opinion.
Request for codes: Never read an unexpected caller a one-time login or 2FA code; call the organization yourself.
Unexpected familiar voice: A recognizable voice on a call you did not expect is not proof of identity.
New number or platform switch: “Text me on this number instead.” Verify through a channel you already trust.

How AI voice scams actually work

The mechanics are simpler than people assume. A scammer needs a short clip of someone’s voice – pulled from a social media video, a podcast guest spot, a voicemail greeting, or a recorded work call. The FBI says even a few seconds is enough to build a convincing clone, and some commercial tools advertise cloning from just three seconds of sample audio. Feed that clip into a voice-cloning model – the same class of fast, cheap AI now driving down the cost of every model release – and the attacker can make “you” say anything in real time.

The cloned voice is only half of it. The other half is social engineering – the script. Criminals pair the voice with a story engineered to bypass your judgment: a grandchild in jail, a CFO authorizing an urgent transfer, a bank “fraud department” needing your code to stop a charge. The voice makes the story believable. The urgency makes you act before you check.

Here is the uncomfortable part: you can no longer reliably catch these by ear. Older advice told you to listen for robotic flatness, odd pauses, or no background noise. Those tells still occasionally show up, but the FBI now warns that AI-generated voices “can sound nearly identical” to the real person, and that the technology has “advanced to the point that it is often difficult to identify.” Listening alone is unreliable. Verification by process is not.

The scale is no longer theoretical

2025 was a record year for online fraud, and AI earned its own chapter for the first time. According to the FBI’s 2025 Internet Crime Report, Americans reported nearly $20.9 billion in cyber-enabled losses across more than one million complaints – the first time complaint volume crossed a million in the center’s 25-year history.

Source: FBI 2025 Internet Crime Report (IC3)
Buried in that report is the number that matters here. For the first time ever, the FBI broke out a dedicated section on artificial intelligence: 22,364 complaints with an AI nexus, totaling roughly $893 million in losses. The bureau added that the real figure is almost certainly higher, because most victims never realize AI was involved in the message, voice, or video used against them. Government-impersonation complaints alone nearly doubled year over year, from about 17,400 in 2024 to more than 32,400 in 2025.

Two faces of the same scam

Voice cloning attacks split cleanly into two categories. The defenses overlap, but knowing which one is aimed at you changes how you respond.

The family emergency call

You get a panicked call. It’s your grandson – he’s been in a wreck, he’s in jail, he needs bail money now, and please don’t tell his parents. The voice is right. The fear is real. The FTC has been warning about this exact scheme since voice cloning went mainstream, and its advice is blunt: don’t trust the voice. Call the person back on a number you know is theirs. If you can’t reach them, contact another family member. Scammers steer you toward gift cards, wire transfers, or crypto precisely because those payments are nearly impossible to claw back.

The fake executive – and why businesses lose the most

In January 2024, a finance employee at the engineering firm Arup joined a video call with the company’s CFO and several colleagues. Everyone on the call looked and sounded familiar. Over the course of the conversation, the employee was directed to make a series of “confidential” transfers – 15 transactions totaling about $25.6 million – to five Hong Kong bank accounts. Every face and voice on that call except the victim’s was an AI deepfake, built from publicly available video and audio of Arup’s real executives.

No system was hacked. No password was stolen. As Arup’s then-CIO Rob Greig put it, this was technology-enhanced social engineering – the attack didn’t go through the firewall, it went through a person. The employee had been suspicious of the initial email; the multi-person video call is what dissolved the doubt. For founders and finance teams, that is the lesson: the weak point is a human under pressure, and the fix is a process that pressure can’t override.

You can’t out-listen a modern voice clone. Verify the request through a channel you control.

The verification protocol that actually works

Forget trying to out-listen the AI. Build habits that don’t depend on detecting the fake at all. These steps come straight from FBI and FTC guidance.

Hang up and call back. End the call and dial the person or organization on a number you already have – never the number that called you, and never a link they send. This check can expose an impersonation when you reach the genuine person.
Use a family code phrase, not a security question. Agree on a secret word with your family in advance. Don’t rely on “what’s your dad’s name?” – that answer is often sitting on someone’s social media. A pre-shared phrase the caller can’t look up is the real test. The FBI explicitly recommends creating a secret word with family members to verify identities.
Refuse the pressure. Legitimate requests survive a pause. Tell the caller you’ll call back, then do it on your own terms. Urgency that can’t tolerate a five-minute delay is the scam itself.
Never share codes or credentials by phone. Never read a one-time passcode or 2FA code to someone who called you unexpectedly. Contact the organization through its official channel.
For any money request, verify out-of-band. Independently confirm wire or payment instructions through a known channel before sending anything – especially if the request is “confidential” or comes from a senior figure.

For businesses: bake it into policy

Individual vigilance isn’t enough inside a company – pressure from a “CEO” overrides instinct. The controls that stop the Arup scenario are structural: dual approval for wire transfers above a set threshold, a mandatory callback protocol for any phone or video payment request, and a culture where a junior employee can pause a transaction the CFO “ordered” without fear. Limit how much executive audio and video sits publicly online, since that footage is exactly what trains the clone.

What the law is doing about it

Regulators have started to move. In February 2024, the FCC ruled that AI-generated voices count as “artificial” under the Telephone Consumer Protection Act – meaning robocalls using cloned voices fall squarely under existing restrictions and give state attorneys general a direct tool to pursue the people behind them. The ruling followed a fake robocall that mimicked President Biden’s voice ahead of the New Hampshire primary. It won’t stop a determined overseas scammer, but it closes a legal gap and signals that synthetic voices are now firmly on regulators’ radar. The same generative leap powering legitimate tools – the kind we cover in our look at how Microsoft and Anthropic stack up on benchmarks – is what makes a three-second clone possible in the first place.

Key facts: AI voice scams in 2026

Commercially advertised audio minimum: as little as 3 seconds; results vary
Total 2025 cyber-fraud losses (US): ~$20.9 billion, a 26% jump and an all-time high
AI-related losses (first time tracked): ~$893 million across 22,364 complaints – likely undercounted
Documented deepfake fraud example: $25.6 million (Arup, 2024)
Best single defense: hang up and call back on a known number
Report a scam: ic3.gov (FBI) and reportfraud.ftc.gov (FTC)

Frequently asked questions

Can I tell if a voice on the phone is AI?

Often not by listening alone. Modern clones can sound nearly identical to the real person. Occasional tells include a slight delay before responses, an unnaturally calm tone, or no background noise – but the FBI warns these are unreliable. AI detection tools have the same problem: when we tested AI text detectors like Turnitin and GPTZero, they proved far less dependable than their marketing claims. Verify through a separate channel instead of trusting a tool or your ear.

How much audio does a scammer need to clone a voice?

Very little. Some commercial services advertise voice cloning from roughly three seconds of clear audio; the quality depends on the sample and tool. That clip can come from a social media video, a voicemail greeting, or a recorded meeting.

Why do scammers ask for gift cards or crypto?

Because those payments are fast and nearly impossible to reverse or trace. Once you hand over gift card numbers or send cryptocurrency, recovery is rare. Any request to pay this way is a major red flag.

What should I do the moment I suspect a scam call?

Stop engaging and hang up. Don’t confirm any personal details, don’t send money, and don’t click links. Then call the supposed person or organization back on a number you already trust to verify the story independently.

Is a family code word really necessary?

It’s one of the most effective low-tech defenses. A pre-agreed secret phrase can’t be found online, unlike answers to “security questions” such as a parent’s name or a pet’s name. The FBI recommends setting one up with your family.

Can I get my money back after a voice cloning scam?

Sometimes, if you act fast. Contact your bank or the payment provider immediately to try to freeze or reverse the transfer, and report it to the FBI at ic3.gov and the FTC at reportfraud.ftc.gov. Recovery odds drop sharply with gift cards and crypto, so speed matters.

Sources and verification

FBI IC3 2025 Internet Crime Report — complaint and loss figures.
FTC advice on AI family emergency scams — callback and payment red flags.
FCC 2024 ruling — artificial voice robocall restrictions.
Original article and interactive red-flags table: Meikuio.

Who Owns Claude? Inside Anthropic — and Why It’s About to Change

Egor Sivchik — Sat, 03 Oct 2026 10:39:46 +0000

Highlights:

The founders and an independent Trust control Anthropic — not Amazon or Google

Anthropic hit $965B and filed for IPO on June 1, 2026, passing OpenAI

A Long-Term Benefit Trust elects a board majority within 4 years

Claude is built by Anthropic, a company controlled by its founders Dario and Daniela Amodei together with an independent trust — not by Amazon or Google, despite the tens of billions those two have poured in. That gap, between who funds Anthropic and who controls it, is the whole story. And as of June 1, 2026, it’s about to be tested in public for the first time.

Anthropic just filed confidentially for an IPO, days after a funding round valued it at $965 billion — higher than OpenAI. So the question people keep typing into Google, “who owns Claude,” is getting a more complicated answer than it had a month ago. Here’s how ownership actually works, who the players are, and why the structure matters more than the dollar figures.

Quick Answer: Who Controls What

The pattern jumps out immediately: the names with the biggest checks sit in the bottom rows, with the least control. That’s not an accident. It’s the result of a corporate structure Anthropic designed on purpose.

The founders: two siblings who walked out of OpenAI

Anthropic was founded in 2021 by Dario Amodei and his sister Daniela Amodei, along with a group of senior researchers who left OpenAI together. Dario had been OpenAI’s VP of Research, leading work on large language models. Daniela had run operations. They didn’t leave over pay or titles — they left over direction, specifically disagreements about how aggressively to prioritize AI safety as the technology scaled.

That origin matters because it’s baked into the company’s legal DNA. Anthropic was built from day one to keep its mission insulated from the usual pressure that investors exert on a fast-growing startup. Dario serves as CEO, Daniela as President, and both sit on the board. They’re joined there by a notable roster: Netflix co-founder Reed Hastings, former Microsoft and GM CFO Chris Liddell, Novartis CEO Vas Narasimhan, and investor Yasmin Razavi.

Money vs. control: why $33 billion doesn’t buy the company

Here’s the part most explainers get wrong. They list the investors, add up the billions, and imply that whoever wrote the biggest check owns Anthropic. That’s not how it works.

Think of it like a mortgage. A bank can lend you most of the money for your house, but the bank doesn’t own your house — you do. Anthropic’s investors are closer to lenders-with-upside than owners-in-control. Amazon has committed enormous sums and is the largest single investor, but it holds a minority equity stake with no board seat and no voting rights. Google has committed even more in cash and cloud compute, but its stake is contractually capped and it, too, has no board seat and no governance control.

Both get something they want — Anthropic runs on their cloud infrastructure — but neither can dictate strategy, product decisions, or who runs the company.

Economic ownership (who holds equity) and operational control (who makes decisions) are two different things at Anthropic, and the company deliberately split them apart. The investors own a slice of the upside. The founders and an independent trust hold the steering wheel.

The investors: who actually put up the money

Anthropic has raised through a long series of rounds, A through H. The most recent — Series H, announced May 28, 2026 — brought in $65 billion at a $965 billion post-money valuation, the largest private valuation of any AI company to date. The lead investors give a sense of who’s betting on Claude at the highest level.

One honest caveat: exact ownership percentages are not public. Anthropic is a private company and doesn’t disclose a cap table. Anyone claiming to know that “Amazon owns 18%” or any precise figure is guessing. What’s confirmed is the direction — large institutional and strategic capital, concentrated among cloud providers and top-tier funds, with founders diluted on paper but still firmly in control through the structure described below.

The money is also tied to compute. Anthropic has signed agreements with Amazon for up to five gigawatts of capacity, with Google and Broadcom for five gigawatts of next-generation TPUs, and with SpaceX for GPU access. AWS remains its primary cloud and training partner. In AI, capital and compute are the same conversation — you raise the first to buy the second.

The Long-Term Benefit Trust: the part nobody else explains

This is the mechanism that makes Anthropic’s ownership genuinely unusual, and it’s the piece most coverage skips because it’s harder to explain than a list of investors.

Anthropic is a Delaware Public Benefit Corporation (PBC). That status alone lets its directors legally balance shareholder returns against a stated public mission — in Anthropic’s case, “the responsible development and maintenance of advanced AI for the long-term benefit of humanity.” But the founders decided a PBC wasn’t enough of a safeguard. So they built something on top of it: the Long-Term Benefit Trust.

The Trust is an independent body of five members chosen for their backgrounds in AI safety, national security, and public policy — and deliberately given no financial stake in Anthropic. When the company closed its Series C, it created a special class of stock, Class T, held only by the Trust. That stock gives the Trust the power to elect and remove board members, and that power phases in over time on funding and time milestones.

The endpoint, by design: within four years, the Trust elects a majority of Anthropic’s board.

Read that again, because it’s the heart of the answer to “who owns Claude.” The body set to control a majority of the board is not an investor, not the founders personally, and not a tech giant. It’s an independent trust whose members are financially disinterested and whose job is to weigh humanity’s long-term interests against shareholder returns when the stakes are highest.

The current Trustees are Neil Buddy Shah, Richard Fontaine, and Mariano-Florentino Cuéllar. The roster has changed over the years — earlier members stepped away for various reasons, including conflicts of interest and new roles — which tells you the structure is a live experiment, not a finished monument. Anthropic itself calls it exactly that: a governance experiment, with built-in failsafes that let stockholders amend the Trust’s powers only by large supermajorities.

The IPO twist: what going public changes

On June 1, 2026, Anthropic confirmed it had confidentially submitted a draft S-1 registration statement to the SEC for a proposed IPO of its common stock. The company was careful with its language: this gives it the option to go public after SEC review, the number of shares and price aren’t set, and the offering depends on market conditions. It is not the same as a public listing — it’s the paperwork that makes one possible.

The context makes it significant. The filing landed just four days after the $65 billion Series H, and that $965 billion valuation pushed Anthropic past OpenAI, valued at roughly $852 billion in its March 2026 round — the first time the Claude maker has been worth more than its biggest rival. Much of that momentum is being driven by demand for Claude in coding and agentic work, the same wave reshaping tools like GitHub Copilot and its 2026 pricing shake-up.

Reports point to a possible listing window later in 2026, though Anthropic hasn’t confirmed timing.

Here’s the question the IPO raises for ownership: what happens to the Trust when public shareholders enter the picture? Anthropic designed the structure specifically so that going public wouldn’t hand control to whoever buys the most stock. The PBC charter and the Trust’s board-election rights are meant to survive a listing. Whether that holds up under the pressure of public markets — quarterly expectations, activist investors, the gravitational pull of share price — is the real thing to watch.

An IPO is exactly the kind of high-stakes moment this structure was built to withstand. We’re about to find out if it does.

Why ownership matters to you

If you use Claude, build on its API, or run it inside a company, the ownership structure isn’t trivia — it’s a signal about how stable and predictable the product is likely to be.

When a tool you depend on is controlled by a single tech giant, its roadmap can swing with that owner’s priorities — a strategy pivot, an acquisition, a decision to fold it into another product. Anthropic’s structure is designed to make that kind of abrupt shift less likely. No single investor can force a direction. The same founding team still runs the company, with legal mechanisms meant to keep the mission steady even as ownership on paper gets more diluted.

For enterprises signing multi-year commitments to Claude, that continuity has real value.

It cuts the other way too. The Trust’s whole point is that it can, in theory, prioritize safety or long-term concerns over short-term commercial wins. For most users that’s reassuring. For an investor expecting a company to maximize returns above all else, it’s a genuine variable — one that’s never really been stress-tested at this scale, and an IPO is where it gets tested.

Data Box

Company: Anthropic, PBC (Delaware Public Benefit Corporation)

Founded: 2021, by Dario and Daniela Amodei

Product: Claude

Latest valuation: $965 billion (Series H, May 28, 2026)

Latest raise: $65 billion, led by Altimeter, Dragoneer, Greenoaks, Sequoia

Run-rate revenue: crossed $47 billion (per Anthropic, May 2026)

Largest investors: Amazon and Google — both minority, neither with board seats or voting control

Governance: founders + Long-Term Benefit Trust; Trust elects a board majority within 4 years via Class T stock

Board: Dario Amodei, Daniela Amodei, Yasmin Razavi, Reed Hastings, Chris Liddell, Vas Narasimhan

IPO status: confidential S-1 filed with the SEC on June 1, 2026; terms not set

Ownership percentages: not publicly disclosed (private company)

Frequently Asked Questions

Who owns Claude?

Claude is owned and built by Anthropic. Anthropic itself is controlled by its founders, Dario and Daniela Amodei, together with an independent Long-Term Benefit Trust that elects a growing share of the board. Investors like Amazon and Google hold equity but no control.

Does Amazon own Anthropic?

No. Amazon is Anthropic’s largest single investor and provides cloud infrastructure through AWS, but it holds only a minority equity stake with no board seat and no voting control. It cannot direct Anthropic’s strategy.

Does Google own Claude?

No. Google has committed significant capital and cloud compute to Anthropic, but its stake is contractually capped and carries no board seat or governance control. Claude remains an independent Anthropic product.

Is Anthropic a public company?

Not yet. As of June 2026, Anthropic is private. It confidentially filed a draft S-1 with the SEC on June 1, 2026, which gives it the option to go public after review — but no shares are being offered and no listing date is confirmed.

Can I buy Anthropic stock?

Not as a retail investor right now. Anthropic is private and the S-1 filing is only a preliminary step. The announcement itself explicitly states it is not an offer to sell securities (we examined how secondary pre-IPO markets treat unlisted labs in how to buy OpenAI stock). If and when an IPO completes, shares would become publicly tradable.

What is the Long-Term Benefit Trust?

It’s an independent body of five financially disinterested members that holds a special class of Anthropic stock (Class T). It has the right to elect and remove board members, phasing in to a board majority within four years. Its purpose is to balance shareholder returns against Anthropic’s public-benefit mission.

Is Anthropic bigger than OpenAI now?

By valuation, yes — for now. Anthropic’s $965 billion Series H valuation in May 2026 surpassed OpenAI’s roughly $852 billion from March 2026, the first time it has been valued higher than its main rival. Valuations shift with each funding round, so this lead may not be permanent.

My Take

The most interesting thing about Anthropic isn’t the $965 billion number — valuations in this industry have stopped feeling real. It’s that the company built a deliberate firewall between the people with the money and the people with the control, and is now walking that firewall straight into the most money-driven environment there is: the public stock market. Plenty of founders say they’ll resist short-term pressure. Anthropic wrote it into a trust and a special class of stock years before it had to.

Whether that holds when quarterly earnings calls and activist investors arrive is the question worth following — not because the answer is obvious, but because almost no one else building frontier AI has structured themselves to even attempt it. If you rely on Claude, that’s the detail that should interest you far more than who led the last round.

About the original author

Dana Pritchard — Marketing Strategist & Tech Journalist

Dana Pritchard is a marketing strategist and tech journalist with 7 years of experience at the intersection of digital marketing and emerging technology. She has worked with growth-stage SaaS companies across the US and Canada, helping teams evaluate and implement AI-driven marketing tools at scale. At meikuio.com, Dana covers the tools and platforms that marketers, founders, and content teams actually use — from AI-powered SEO platforms to social media automation and analytics.

Her writing focuses on practical outcomes: what moves the needle, what doesn’t, and what the price tag really means for a small team. Dana is based in Austin, Texas, and has been featured in several B2B marketing publications focused on MarTech and demand generation.

Original article: MEIKUIO — Who Owns Claude? Inside Anthropic — and Why It’s About to Change