You are building a slide that says which AI model enterprises actually use, and you need one number you can defend in the room. So you search for Claude AI statistics and get a dozen pages promising forty or fifty data points each. Revenue run rates. User counts. Valuations. Market share figures that disagree with each other by a factor of two.
Then you click through to check one, and the citation points at another blog, which points at a third, which cites nobody.
That is the state of Claude AI statistics as a category, and it is worth knowing before you put any of it in front of a client. The useful news is that one properly documented dataset does exist, and it answers the question most people are actually asking.
The One Dataset With a Methodology
Menlo Ventures publishes an annual enterprise survey, and the 2025 edition is the rare source in AI market research that tells you how it got its numbers.
The method is stated plainly: 495 United States enterprise AI decision-makers, surveyed with an independent research firm between November 7 and 25, 2025, covering C-suite executives, VPs of engineering and product, and technical leaders at companies actively using AI tools. Market sizing combined that survey with a bottom-up model. The authors also note the limitation, which is that the sample covers US enterprises only.
The headline figure for Anthropic is 40% of enterprise LLM API market share. That is up from 24% in 2024 and 12% in 2023, which is the more interesting fact, because the trajectory is steeper than the level. OpenAI sat at 27% in the same data, down from 50% in 2023. Google was at 21%, up from 7%. Those three account for 88% of enterprise LLM API usage, with the remaining 12% split between Meta’s Llama, Cohere, Mistral, and smaller providers.
For context on the pot being divided, the report puts foundation model APIs at about $12.5 billion within an $18 billion infrastructure layer, inside roughly $37 billion of total enterprise AI investment in 2025, up from $11.5 billion the year before. That total split roughly evenly, $19 billion into applications and $18 billion into infrastructure, which is itself a useful corrective to the assumption that enterprise AI money is mostly going to model providers.
Why does the methodology matter this much? Because it is the only thing separating a statistic from a rumour, and in this category almost nobody provides it. A survey with a stated sample size, a stated window, a stated respondent profile, and an acknowledged limitation can be argued with. You can say 495 respondents is thin for a market this size, or that a November window predates three model releases, and those are real criticisms. A figure with no method attached cannot be criticised, which people mistake for strength.
Where Claude’s Lead Is Widest
The aggregate share figure undersells how concentrated the lead is, and the concentration is the part that matters strategically.
Coding is the largest departmental category of AI spend in the report at roughly $4.0 billion, about 55% of all departmental AI spending. Code completion alone reached $2.3 billion, against a cited figure of around $550 million for AI coding spend in 2024. That is a category that grew roughly sevenfold in a year.
Within it, Menlo put Anthropic at an estimated 54% of the coding LLM market against 21% for OpenAI. So the single largest pool of enterprise AI money is also the pool where Claude’s share is highest, which explains the aggregate number better than any general statement about model quality.
Two supporting figures are worth keeping. Fifty percent of developers use AI coding tools daily, rising to 65% in top-quartile organisations, and teams report velocity gains above 15%. Those are adoption and outcome numbers from the same survey, which makes them more useful than the vendor-published productivity claims that circulate alongside them.
One sobering figure sits in the same report. Only 16% of enterprise deployments and 27% of startup deployments qualify as true agents, with agent platforms capturing about $750 million. Most of what gets described as agentic in 2026 is, by this definition, not.
That gap between the language and the deployments is worth holding onto, because it is the single most useful thing in the dataset for anyone selling into this market. The vocabulary has run several years ahead of the implementations. A buyer who says they are building agents is, five times out of six, building a workflow with a model call in it, and a pitch that assumes otherwise lands badly.
It also explains why coding leads the spending table so decisively. Code completion is the one application where the model’s output is immediately verifiable by the person receiving it, the feedback loop is seconds long, and the failure mode is a rejected suggestion rather than a bad business decision. Categories without those properties are still mostly in pilot, which is where the 16% comes from.
Why Do the User Numbers Vary So Much?
Because they measure a thing nobody has published, and that produces a specific failure mode worth naming.
Anthropic does not release a monthly active user count. Neither do most of its competitors in a form that survives scrutiny. So when a post gives you a Claude user figure, it was either estimated from third-party web traffic panels, inferred from app store data, or invented. None of the three is labelled as such, and the number arrives looking identical to the sourced ones.
Here is the distinction I use to keep this straight, and I think of it as the Spend-versus-Seats problem. Enterprise API spend share and consumer user share measure different markets and can point in opposite directions. A model can hold a commanding share of what companies pay for programmatic access while holding a modest share of people typing into a chat box, because the first is bought by engineering leaders optimising for reliability and the second is chosen by individuals optimising for familiarity.
Both are real. Neither substitutes for the other. The error that gets people caught out in meetings is citing a spend-share figure to support a claim about popularity, or a traffic-panel estimate to support a claim about enterprise adoption. State which market your number describes, every time, and most of the apparent contradictions in this category dissolve.
There is a third market people blur into the first two, which is developer tooling. Usage inside a coding tool may be billed as API consumption, counted as seats in a subscription, or bundled inside an IDE licence, and which of those it is depends on the vendor rather than on anything about the user. So a figure describing coding usage can be a spend measure or a seat measure depending on who assembled it, and the posts quoting it almost never say.
The practical rule is short. Before you repeat any share figure, answer two questions: share of what pool, measured how. If the source does not let you answer both, you do not have a statistic, you have a sentence.
Three Numbers You Should Stop Repeating
I traced the most-cited Claude AI statistics outside the Menlo data. Three keep appearing and none of them survives a source check.
Any monthly active user count. Whatever figure you have seen, it is not a disclosure, because there is nothing to disclose from. The posts carrying it do not say whether it means registered accounts, app installs, or unique visitors, which is the tell. Drop it entirely rather than hedging it.
Any revenue run-rate figure that arrives without a named reporting outlet. Private company revenue reaches the public through either a company statement or reporting that cites people with knowledge of the numbers. Several of the run-rate figures circulating for 2026 have neither, and at least one I found is implausible enough on its face that repeating it would cost you credibility with anyone who follows the sector.
And the misquote of the good number, which is the most common error of all. The 40% figure is enterprise LLM API market share. It is not 40% of the AI market, not 40% of AI revenue, and not 40% of AI usage. I have seen all three versions in decks. The original is a strong, defensible statistic and the paraphrase is wrong, which is a bad trade.
The reason that one matters more than the other two is that it is the error a knowledgeable person in the room will catch. Nobody can disprove an invented user count on the spot. Anyone who has read the Menlo report can tell you that enterprise API spend is a narrow slice of a $37 billion market, and the correction lands in front of your client.
Cite it with its qualifier attached, every time, even though the qualifier makes the sentence clunkier. Forty percent of enterprise LLM API spend, per Menlo Ventures’ late-2025 survey of 495 US enterprise decision-makers, is a mouthful and it is bulletproof.
What the Enterprise Share Does Not Tell You
Worth being clear about the limits of the figure I have just spent four sections endorsing.
It is a US enterprise sample, stated as such by its authors. It tells you nothing reliable about Europe, India, or Asia, where procurement patterns, data residency rules, and local model options differ substantially.
It measures spend rather than satisfaction or output. A company can spend heavily on a model it is actively evaluating alternatives to, and migration costs mean spend share lags sentiment by quarters.
It is a point-in-time measurement from late 2025, which in this market is a long time. The 12% to 24% to 40% progression tells you the direction was sharply upward through that window. It does not license you to extrapolate a 2026 figure, and the number of posts quietly presenting an extrapolation as a measurement is the reason this whole article exists.
What This Means for Being Cited by Claude
If you are reading this because you want your business to appear when someone asks Claude for recommendations, the market share data matters for one reason only: it tells you how much enterprise attention flows through this particular model, which is how you decide whether to care.
Forty percent of enterprise API spend and a majority of coding usage means a great deal of professional, high-intent attention passes through it. The people asking are often technical buyers with budget. That makes being named a commercially significant outcome rather than a vanity one.
What it does not change is the work. Being cited by Claude, ChatGPT, or Google’s AI Mode comes down to the same inputs in each case. A clear claim that is specific enough to be useful. A date, so the system knows how current it is. Corroboration from independent indexed sources that describe your entity the same way. Across the 1,016 publication catalogue Instant Press works from, 96% are Google-indexed with mean Domain Authority of 62.3, and for citation purposes the indexing matters far more than the masthead, because these systems are reconciling agreeing sources rather than ranking prestige.
There is no Claude-specific optimisation worth building. There is only being a source worth citing, and then measuring which engines noticed.
The one adjustment the enterprise skew does justify is in what you publish rather than how. If a disproportionate share of the attention flowing through this model belongs to technical buyers and engineering leaders, then the pages most likely to earn a citation are the ones that answer technical questions precisely: documentation, methodology, benchmarks you actually ran, pricing you actually publish. Thought leadership about transformation does not get cited by anything, because there is no claim in it to cite.
Write the page that contains the number someone needs, and put the date on it.
What You Should Actually Track
Three numbers, all of which you can produce yourself and none of which depend on anybody’s market research.
Citation rate against a fixed question set. Write twenty questions your buyers genuinely ask, run them monthly across the engines you care about, and log whether any page of yours is cited. That is a proprietary dataset after one quarter and it measures the outcome directly.
Share of cited sources. When your category comes up, which domains get named? This tells you who the models currently trust in your space, and it is a far better content brief than keyword volume.
Corroboration depth. How many independent indexed sources describe your entity consistently? This is the input variable and the only one you control. It is also the one that moves the other two.
Where This Goes Next
The thing worth watching is not whether Anthropic’s share goes to 45% or back to 30%, because that number will be restated by someone every quarter and most of the restatements will be guesses.
What will matter is the separation already visible in the Menlo data between models bought programmatically by engineering teams and models chosen by individuals in a browser. Those two markets are being measured as one thing and they are diverging, which means the single-number market share framing that dominates every statistics roundup is going to get less useful rather than more. The firms that plan well from here will be the ones tracking their own citation rate in the specific places their buyers ask questions, and treating everyone else’s headline figures as weather rather than as navigation.