Brazil ranks 5th in the world in Claude usage — and that means less than it seems

Brazil is the fifth country in the world by Claude usage. The figure does not come from a consultancy, a scraping panel or "people familiar with the matter": it comes from the microdata Anthropic itself publishes, by country, month by month, under a CC-BY licence, with a replication notebook. I downloaded the dataset and measured it.

The same measurement says a second thing, which no headline carries: in per-capita intensity of use, Brazil is 61st out of 121 countries. We are at the top because we are big, not because we are intense. That distinction — volume is not intensity — is the key to this entire article, because its money version (run rate is not revenue) is the confusion propping up nearly every "Claude statistics" page published in 2026.

I studied the most complete of those pages — a German compilation from April 2026, with dozens of numbers on revenue, valuation, users and benchmarks — and checked every source I could open: the primary ones plus 21 community links, one by one. What follows is the result of that check, including what it turned up against me.

The only primary data on Claude usage is the one nobody cites

The German page admits it, honestly: "unlike OpenAI, Anthropic does not publish detailed user numbers". It is true — and the consequence is that none of the headline numbers comes from the company. Each one has a different producer, with a different method:

The last row is the most revealing: the page credits to a Substack the only number in the table the company actually published. They cited the echo instead of the source. (Claude Code's product numbers — models, usage, autonomy — are left for an article of their own.)

Meanwhile, there is one figure Anthropic does publish: the Anthropic Economic Index, with usage microdata by country, open methodology, a downloadable dataset on Hugging Face under CC-BY and replication notebooks. It is the only figure on the list anyone can reproduce — and the page cites it once, in passing, for a Singapore number. The only primary, documented figure gets one line; the leaked run rate gets the headline.

The cost of ignoring the primary source has a numerical example. An American aggregator claims "245 million monthly active users as of mid-2026", crediting a Sensor Tower report that is paid and closed — I could not confirm the report contains the number. Another, using SimilarWeb, said "~20 million" in September 2025. A twelvefold gap, neither primary, and a suspicious pattern in the larger aggregator: 245 million users x US$ 192 per user ≈ US$ 47 billion of "ARR" — the three numbers close perfectly against each other, which suggests two of them were derived from the third, worked backwards.

5th in volume, 61st in intensity: Brazil uses Claude in proportion to its size

I downloaded the dataset from the Economic Index's June 2026 release (219 MB of CSV) and measured Brazil with a 73-line script. The numbers for the month of May 2026:

The AUI (Anthropic AI Usage Index) is a ratio: Brazil's share of Claude usage divided by Brazil's share of the world's working-age population. AUI = 1 means using exactly in proportion to your own size. Brazil scores 0.96 — and that is what the 5th place does not tell you: we are among the world's largest volumes because we are a country of 200 million people, not because the average Brazilian uses Claude more than the rest of the world does.

The top 10 by volume carries the entire spectrum of that distinction:

India is the world's 2nd-largest volume with an AUI of 0.30 (104th of 121); Australia is the 9th volume with an AUI of 6.40 — 1st in the world in intensity. Volume measures size; AUI measures behaviour. They are different rulers, and a ranking that shows only one of them is hiding the other half of the information.

Three caveats any such comparison requires — and that statistics pages never make:

1. The denominator is not 100%. The 121 countries published in the 2026 release add up to 87.5% of global usage; the rest has no publicly attributed country. In the 2025 release the hole was explicit: a not_classified bucket holding 15.66% of all usage.
2. Rank across releases is not a clean comparison. The two releases have different coverage (194 vs 121 countries) and different windows (one week of August 2025 vs the full month of May 2026). Brazil's "3rd -> 5th" says less than it seems. The AUI does compare — it is a ratio against population, immune to how many countries get published.
3. The stable measure is intensity. In ten months Brazil's AUI went from 0.93 to 0.96 — practically flat, and it is the most stable measure in the whole set. What shifts position is the ranking around it, not Brazilian behaviour.

And what do Brazilians use it for? 57.4% of usage is declared as work — according to Anthropic's own geography report, Brazil and the Balkans have the highest proportion of professional usage in the world. The microdata confirms the prose:

Methodological note. Every number in this section comes out of the open Anthropic/EconomicIndex dataset (June 26, 2026 release), measured by a 73-line script that accompanies this article — the same ruler at both ends, never a prose summary checked against microdata. The AUI uses working-age population (15-64) as its denominator, per the definition Anthropic publishes. Reproducing it takes two commands: download the CSV and run the script.

Run rate is not revenue — and the page knows it

The German page defines run rate correctly: annual revenue extrapolated from the current month. Two sections later, it divides that run rate by headcount and calls the result "revenue per employee" (US$ 4.2 million per head), to conclude that Anthropic is "far more efficient than the traditional giants". Further down, it divides valuation by run rate (20x vs OpenAI's 29x) and concludes that Anthropic "operates more efficiently" — a valuation multiple does not measure operational efficiency; it measures what investors pay per unit of extrapolated revenue. And the denominator mixes rulers: Anthropic's US$ 380 billion was post-money; OpenAI's US$ 730 billion, as the ChatGPT statistics article showed, was pre-money. The calculation compares an orange with a peeled orange.

What "run rate" means in practice, nobody has explained better than Reuters Breakingviews (via Simon Willison): for API consumption, Anthropic takes the last 28 days and multiplies by 13; for subscriptions, the current month times 12. "Annualised" is not the year — it is the present projected as if growth stopped today, at a company growing more than 10x a year.

The 2026 run rate timeline, with the ruler of who said each number:

The sequence is real: US$ 14 billion signed by the company on February 12; "nears $20 billion" via Bloomberg on March 3 ("recently surpassed $19 billion… up from $9 billion at the end of 2025"); US$ 30 billion announced on April 6, per Willison's timeline; US$ 47 billion on the Series H page, in May. None of it is a lie — Willison is blunt: lying to investors who had just put in US$ 65 billion would be securities fraud.

But in the middle of that sequence sits a number of a different nature. On March 9, 2026, in Anthropic's case against the U.S. Department of War (case 3:26-cv-01996, N.D. California), CFO Krishna Rao declared, under penalty of perjury: the company has generated revenue "exceeding $5 billion to date" — since its founding. In the same window in which the public run rate was ~US$ 19-20 billion. The same filing records more than US$ 10 billion already spent on training and inference, and more than US$ 60 billion raised in outside capital.

That is not a contradiction — it is the definition. Davi Ottenheimer, who analysed the filing at flyingpenguin, gave the exact image: run rate is the speedometer; cumulative revenue is the odometer. At a company growing 10x a year, the speedometer reading 19 billion a year with the odometer at 5 billion is arithmetic, not scandal: nearly all the revenue in the company's history happened in the last few months. Whoever publishes the ">US$ 5 billion" as "they caught Anthropic lying" is committing exactly the error this article describes — swapping a measure of accumulation for a measure of pace.

The best the sceptics have is not scandal; it is a ruler. Jason Lemkin (SaaStr): the US$ 47 billion is a strong month annualised — calendar-2026 revenue should land at US$ 20-26 billion, top five in software, not number two. Ottenheimer points out that the public curve implies ~US$ 6.7 billion accumulated through March against the sworn ">5 billion" — a gap that could be 25-35% overstatement or simply ARR counting contracts signed before the money comes in. And Ed Zitron, the harshest of the critics, circulated the CFO's declaration with the wrong date — March 6, when the filing was executed on the 9th — a small error, but the kind of error this fact-check exists to catch: I cited the filing, not the blog that summarised it.

When "32%" means "150 people answered"

The enterprise market share is the German page's most-cited number: "Anthropic 32%, OpenAI 25%". The real chain of attribution: the page credits TechCrunch, which only reported it; the number is Menlo Ventures' — and it is an opinion survey. The edition used (Mid-Year 2025) heard from 150 technical leaders. "32% share" means: out of 150 people who answered a form, a third said they spend more with Anthropic.

Worse: by the time the page was published, in April 2026, Menlo had already released the new edition four months earlier (495 respondents, November 2025): Anthropic at 40% of enterprise LLM spend, OpenAI 27%, Google 21% — and, in coding, Anthropic at 54% against the runner-up's 21%. The page used the stale edition of the source it presented as measured fact.

And the page's chart makes an error neither source made: it merges two incompatible series. Ramp's (corporate-card spend by Ramp's customers: Anthropic from ~10% to >65% of combined spend) and Menlo's (a survey about the API market: from 12% to 32%) — with the same "OpenAI: from 50% to 25%" pair appearing in both, over different time windows, credited to different sources. A corporate card measures the spending of whoever uses Ramp; a survey measures the answers of whoever got asked. Stacking the two into one chart only produces a curve no instrument measured.

The practical rule: a survey introduces itself with its N. "40% (survey of 495 American decision-makers, self-reported)" is information; "40% of the market" is precision fiction.

The right table, read in the wrong order — the mistake I made too

The page's only mention of the Economic Index says: the highest per-capita usage is in "Singapore (4.6x above average), followed by Canada (2.9x), Israel and Australia". I measured it in the dataset of the release it cites (2025, 194 countries):

All four values on the page exist and check out. The defect is one of ordering: Singapore appears as the highest (it is 3rd), Canada as second (it is 12th), and Israel — 1st in the world, by a wide margin — is cited third, with no number. They read the right table and reordered it wrong.

I record what the check turned up against me: my first reading of that passage concluded that Canada's "2.9x" had been made up, because Canada did not appear in my top 5. That reading was wrong. The value exists — 2.914, exactly the "2.9x" — what does not exist is the position the page gives it. I only found out by measuring the dataset instead of trusting my own re-reading. The ordering error is the easiest one to commit because the shallow check — "do the numbers exist?" — passes.

Two conceptual defects complete the picture. "4.6x above average" misdescribes the instrument: AUI is not deviation from a mean, it is a ratio against population (Singapore uses 4.6x the proportion of its own size, not "4.6x the average"). And the top of the AUI ranking is noisy by construction: Monaco, Luxembourg, Iceland and Malta show up there because they are tiny — in a small population, few people move the ratio a lot. Reading the AUI as a "ranking of the most AI-advanced countries" is misreading the instrument — it measures relative intensity, with all the noise small denominators bring.

What you can check today with no intermediary

To be fair to the page: a good part of what it publishes matches the primary source. Prices, above all — checked against the official table in August 2026:

The problem shows up in what the page claims against itself:

1. The context window has three answers on the same page. 200K tokens in two sections, "1M standard" in two others, "1M in beta" in a third. Anthropic's documentation, at the time, gave 1M for Opus 4.6 and Sonnet 4.6 — two of the three answers were wrong, and the reader has no way of knowing which.
2. A sentence that contradicts itself on its own: "GPT-5.4 is far more expensive on output (US$ 15, against Sonnet 4.6's US$ 15)". Fifteen against fifteen.
3. The decimal benchmark the source does not publish. The page compares Opus 4.6 (80.8% on SWE-bench) with Opus 4.5 (80.9%) — the newer model scoring lower, inside a narrative of progression. I went to the primary sources to reproduce the pair and could not: the benchmark tables in both official announcements are images, and the only textual number is a footnote on the 4.6 — 81.42%, with prompt modification. The page compares to a tenth of a point what the source does not publish as a comparable pair.
4. "#1 on all three LMSYS leaderboards at the same time" — Anthropic never claimed that. Zero mentions of Arena or leaderboard in the Opus 4.6 announcement. The leaderboard's April snapshot is not accessible; I cannot prove the claim was false on that date — I can state it was not Anthropic's and that today it is not true.
5. The knowledge-cutoff table mixes two definitions. Anthropic publishes two cutoffs per model — reliable knowledge and training data — and the page uses now one, now the other, in the same table: Sonnet 4.5 with the reliable one, Opus 4.6 with the training one, Opus 4.5 with a date that is neither.

The pattern across the five failures is the same: all of them are checkable today, for free, with no intermediary — in the documentation and the public announcements. What requires a leak, the page handles with care; what only requires opening the source, nobody opened.

A four-month half-life

The page is dated April 3, 2026. Between April and August:

The Claude Code episode earns its place for the chain of attribution, which the community harvest had wrong: the removal was reported by a user, George Pu; Anthropic's head of growth, Amol Avasare, published only the walk-back — "small test on ~2% of new prosumer signups. Existing Pro and Max subscribers aren't affected". Even the product's perimeter changes on a scale of weeks, and by A/B test.

None of this is the German author's fault — in April, April's numbers were the numbers. The fault is the genre's: a "2026 statistics" page has a half-life of weeks, and the format — dozens of numbers, a single date at the top — ages as a block and in silence. The July reader reads March numbers believing they are reading the present.

What I would do with this

Ask who signed the number before asking what the number is. The same quantity — "Anthropic's revenue" — circulated in 2026 as US$ 14 billion (the company, in an announcement), ~US$ 19 billion (Bloomberg, via leak), US$ 47 billion (the company, a strong month's run rate) and >US$ 5 billion (the CFO, under oath, measuring something else). None is a lie; each answers a different question. Whoever does not know which question a number answers does not know what the number says.

Separate the speedometer from the odometer in every revenue figure. Run rate is instantaneous pace, projected; revenue is what came in. The distance between the two is proportional to growth — and at a company growing 10x a year, it is enormous and normal. Distrust whoever presents the distance as a smoking gun, and whoever presents the speedometer as an odometer.

If the subject is Claude usage, go to the only open data there is. The Economic Index is CC-BY, has a replication notebook and microdata by country. Every Brazilian number in this article comes out of a 73-line script over a public CSV — anyone can reproduce it in two commands. That is more than can be said of any other statistic cited here, and it is why Brazil-5th-in-volume-61st-in-intensity is the most solid number in this text.

And the next time a statistics page shows up in your feed, the cheap test is this: find one claim that is free to check — a price, a context window — and check it. If the author did not open the source that costs nothing, the ruler for everything else is set.

Sources

• Primary usage data: Anthropic Economic Index · dataset on Hugging Face (CC-BY, June 26, 2026 release)
• Anthropic announcements: Series G, US$ 30bn / US$ 380bn post-money, Feb 12, 2026 · Series H, US$ 65bn / US$ 965bn post-money, May 2026
• Declaration by CFO Krishna Rao (Mar 9, 2026), case 3:26-cv-01996 (N.D. Cal.), Anthropic PBC v. U.S. Department of War: CourtListener, Document 6-5
• Run rate timeline and mechanics: Simon Willison, May 29 · May 31
• Press: Bloomberg, Mar 3, 2026 · TechCrunch, Mar 28, 2026
• Enterprise survey: Menlo Ventures, State of Generative AI in the Enterprise, Dec 2025
• Sceptics and rulers: Jason Lemkin / SaaStr · Davi Ottenheimer / flyingpenguin, Mar 17, 2026 · Ed Zitron, May 26, 2026 · David Gerard / Pivot to AI, Apr 23, 2026
• Page analysed: gradually.ai/claude-statistiken (Wayback snapshot, Apr 18, 2026)

Verification: Brazil's numbers and the AUI rankings measured by the author on the Economic Index's open dataset (June 26, 2026 release) in August 2026; prices checked at claude.com/pricing on Aug 3, 2026; announcements, posts and the court filing read in full between Aug 3 and 10, 2026 — Bloomberg and Pivot to AI via Wayback Machine snapshots; the two tweets cited were checked by direct navigation. I did not read the Wall Street Journal article cited by the harvest (paywalled even on the Wayback) — nothing here depends on it. I could not verify the April snapshot of the LMSYS leaderboard; the text marks that claim as non-reproducible, not as false. The CFO's filing was cited from the court document, not from third-party summaries.