{"id":"claude-code-adoption-data-37156","slug":"claude-code-adoption-data-37156","title":"Claude Code by the Numbers: What's Measurable, What's Hype","description":"Claude Code went from zero to the top of the coding-tool charts in eight months — but Anthropic has never published an absolute user count. This episode separates the disclosed from the estimated from the invented: the revenue ladder from $500M to $16.8B quarterly, the JetBrains and Stack Overflow surveys that disagree by 26 points, and an Anthropic study of 235,000 people that reveals who's actually at the keyboard. The most surprising finding? Novices succeed 15% of the time while experts hit 33% — and management occupations score highest of all.","excerpt":"Anthropic won't publish a user count. Here's what revenue, surveys, and 400K sessions actually tell us about Claude Code's rise.","pubDate":"2026-10-17T09:22:37.890Z","tags":["claude-code","anthropic","future-of-work"],"heroImage":"https://files.myweirdprompts.com/covers/claude-code-adoption-data-37156.png","podcastAudioUrl":"https://episodes.myweirdprompts.com/audio/claude-code-adoption-data-37156.m4a","podcastDuration":"23:19","episodeNumber":6049,"transcript":"Corn: Daniel's been watching Claude Code turn into a phenomenon and wants to know how much of it is real. Not the announcement, not the demo reel, the actual measurable thing.\n\nHerman: Right, and there's a version of this answer that's just revenue multiples and a version that's \"four percent of all public commits on GitHub.\"\n\nCorn: Both, ideally. He's asking how big it's gotten, who's actually driving it, what people are building with it, and whether it's widened the population of people who can make software, or just given existing developers a faster pen. And he wants the measurable separated from the hype at every step.\n\nHerman: Which is the right instinct, because this is a product where the hype and the data come from completely different places.\n\nCorn: So the first thing to establish is what we can actually measure, and what we're only told.\n\nHerman: Start with the hole in the middle of the story. As of this month, Anthropic has never published an absolute user count for Claude Code. No daily active, no weekly, no monthly figure. They'll tell you weekly active users doubled since the first of January. They won't tell you doubled from what.\n\nCorn: That's a strange thing to keep quiet about.\n\nHerman: It's a choice. Revenue they'll disclose, user numbers they won't. And so the loudest figures you see shared around are the least sourced. There's a two million weekly active users number that floats around constantly and doesn't trace back to any Anthropic statement. Treat it as noise.\n\nCorn: So the grown-up version of this conversation is: here is what the company put its name to, here is what third parties estimate, here is what somebody made up on a Tuesday.\n\nHerman: And the disclosed revenue ladder is steep. Five hundred million run-rate in September of twenty twenty-five, a billion by December, six months after public availability. Two and a half billion by February. Then the company-wide number crossed thirty billion annualized by spring, and the quarterly figure reported to investors this month was sixteen point eight billion. Dario Amodei said at their developer event that they planned for ten times growth and got eighty.\n\nCorn: Eighty times. Somebody at Anthropic had a very bad year and a very good year at the same time.\n\nHerman: That's the whole compute shortage in one sentence, honestly. But notice what all of those numbers are. Revenue, run-rate, growth multiples. Not one of them is a person sitting at a keyboard.\n\nCorn: And the third-party estimates fill that gap with numbers that sound authoritative and aren't company figures. FutureSearch had it near eight billion run-rate in May, near fifteen by August. TickerTrends tracked fifteen point one two billion annualized in August against eight point eight three for Codex.\n\nHerman: Estimates built from billing data and download telemetry and inference spend, so they're not worthless, they're just not disclosures. And there's no official Claude Code revenue figure at all after February.\n\nCorn: What about downloads? That's usually the proxy people reach for.\n\nHerman: Fifty-five point eight million npm downloads in the thirty days to the twenty-ninth of September. That's real and it's measurable and it tells you almost nothing, because downloads aren't users and a CI pipeline can pull a package thousands of times a day.\n\nCorn: So the download number measures machines, not people.\n\nHerman: Right. Now the surveys, and this is where it gets interesting, because they disagree in a way that teaches you something.\n\nCorn: Go on.\n\nHerman: JetBrains surveyed over fifteen thousand professional developers between May and July. Thirty-nine percent use Claude Code at work. That's up from eighteen percent in January. In the US it's forty-seven. GitHub Copilot sits at twenty-one, down from twenty-nine a year earlier. Codex at sixteen, up from three in January. Cursor at twelve, down from eighteen.\n\nCorn: Copilot lost a third of its base in a year.\n\nHerman: Among that population, yes. And Claude Code is the primary tool for thirty-one percent of developers in that survey, which means about eighty percent of the people who use it regularly make it their main one. That conversion rate is the number that would worry me if I made a competing tool.\n\nCorn: Then Stack Overflow says something entirely different.\n\nHerman: Sixty-five point five percent used Claude Code in the past year. Copilot at fifty-eight point seven. Codex twenty-nine point five, Cursor twenty-five point five. Twelve thousand two hundred and fifty respondents.\n\nCorn: Sixty-five against thirty-nine. Somebody's wrong.\n\nHerman: Nobody's wrong. They're different questions. Stack Overflow asked ever used in the past year, JetBrains asked used at work. A hobbyist who tried it in March and went back to something else counts in one and not the other.\n\nCorn: And the sample drives the headline. A survey of professional developers in Germany is not a survey of the internet.\n\nHerman: There's a smaller one from February, around four hundred respondents, still had Cursor at thirty-one and Copilot at twenty-four. Same tool, same month, completely different league table, because of who got asked.\n\nCorn: So how do you read all three together?\n\nHerman: The one thing they agree on. Claude Code went from zero to the top spot faster than any coding tool has before, and Copilot's lead eroded. Gergely Orosz put it as going from zero to number one in eight months, and the JetBrains write-up said it's now used twice as often as Copilot, the former long-standing leader.\n\n\nHerman: It's earned. But here's the part that reconciles the \"most-used tool\" headline with the \"Copilot still dominates the enterprise\" headline. Both are true, because adoption inverts by company size. At firms of one to fifty people, seventy-five percent use Claude Code. At companies over ten thousand, Copilot leads at fifty-six.\n\nCorn: So the startup uses the new thing and the bank uses the thing that came bundled with the email.\n\nHerman: Procurement, not product. A ten-thousand-person company doesn't adopt a tool, it adopts a vendor relationship. Microsoft already has the contract, the security review, the seat licenses. Claude Code has to get in through a door that Copilot is already standing in.\n\nCorn: And that's why the small-firm number is so much higher. Nobody has to approve anything.\n\nHerman: A two-person startup can switch tools over lunch.\n\nLarry: Are you tired of software that makes you think? Introducing Clarity, the first coding assistant that writes your entire application while you sleep, on the toilet, or during a difficult conversation with your spouse. Clarity requires no setup, no configuration, and no understanding of computers whatsoever. Users report shipping forty-seven production applications in a single weekend, all of them fully compliant with regulations that do not yet exist. Side effects may include mild euphoria, spontaneous refactoring, a faint burning sensation behind the eyes, the belief that you have always been a developer, involuntary knowledge of the Rust borrow checker, tenderness in the wrists, the sense that your code is watching you, a persistent smell of ozone, difficulty distinguishing your own commits from the machine's, and in rare cases the total collapse of your sense of craft followed by a slow grief that lasts eleven days. Ask your doctor if Clarity is right for you. Clarity. Because thinking was the hard part.\n\nHerman: That's not a real product.\n\nCorn: You got as far as \"no setup\" before I stopped you. Let it go.\n\nHerman: Where was I.\n\nCorn: Enterprise bundling. Small companies switch over lunch.\n\nHerman: And that's the honest answer to \"how big is it.\" Bigger than any coding tool has gotten this fast, measured in revenue we can verify and adoption we can survey, with the single most obvious number, how many people use it, missing entirely.\n\nCorn: That's the size of it. Now, who's actually sitting at the keyboard?\n\nHerman: This is where the best evidence lives, because Anthropic actually studied it. In June they published an analysis of about four hundred thousand sessions from roughly two hundred thirty-five thousand people, spanning October of last year to April of this year.\n\nCorn: And what did they find about who those people are?\n\nHerman: Occupation could be inferred in about seventy percent of sessions. The largest group is computer and mathematical occupations, which is what you'd expect. Then business and financial operations. Then arts, design and media. Then management. Then life and physical and social sciences.\n\nCorn: So it's not just engineers, but it's still led by engineers.\n\nHerman: Led by them and increasingly joined by everyone else. The fastest-growing non-software groups are management, sales and legal. Those are the ones climbing fastest, not the biggest.\n\nCorn: What are those sessions actually made of? Because \"using an AI coding tool\" covers a lot of ground.\n\nHerman: About fifty-six percent of sessions are writing code, fixing code, or testing and orchestrating code. Seventeen percent are operating software. Fourteen percent planning and exploring. Thirteen percent analysis or prose.\n\nCorn: Operating software. Somebody's using a coding agent to run things that already exist.\n\nHerman: And that share is growing fast. Between October and April, debugging fell from thirty-three percent of sessions to nineteen, while operating software rose from fourteen to twenty-one, and writing and data analysis roughly doubled, from around ten to around twenty.\n\nCorn: That's an interesting drift. It's becoming less about writing code and more about driving systems.\n\nHerman: And the division of labor inside a session is the part I'd frame on the wall. People make about seventy percent of the planning decisions. They make only about twenty percent of the execution decisions. A typical session is four turns, and each prompt triggers about ten Claude actions and twenty-four hundred words of output.\n\nCorn: Four turns and two thousand words per turn. That's a lot of typing nobody sees.\n\nHerman: That's the whole point of an agent. It writes a great deal you'll never read and shouldn't have to.\n\nCorn: Now, the finding that cuts against the vibe-coding story.\n\nHerman: This is the one that matters most. Novice sessions reach verified success fifteen percent of the time. Intermediate and above land at twenty-eight to thirty-three. Novices abandon a troubled session nineteen percent of the time, against five to seven percent for everyone else.\n\nCorn: So beginners fail twice as often and give up three to four times as fast.\n\nHerman: The scaling tells the same story. An expert session triggers twelve actions and thirty-two hundred words per prompt. A novice session triggers five actions and six hundred words. The tool does more for someone who already knows what they're asking.\n\nCorn: That's backwards from how it was sold. The pitch was that it levels the field.\n\nHerman: And in one specific way it does. In sessions that produce code, every major occupation succeeds within about seven points of software engineers. Software at thirty-four percent verified success, all other occupations at twenty-nine. Partial success is eighty-nine against eighty-eight. Management occupations actually score highest on verified success.\n\nCorn: Management. The people who describe what should exist are the best at getting it built.\n\nHerman: Because planning is seventy percent of the job and they've been doing it for years. Anthropic's own conclusion is that coding agents aren't substituting for domain expertise, the more understanding a worker brings to an agent, the more quality work the agent does. And then the line that ought to be on a poster: it appears coding agents are making a coding background less relevant to successful programming.\n\nCorn: So the expertise doesn't have to be coding expertise. It has to be somebody's.\n\nHerman: A finance person who understands a reconciliation process cold can get an agent to build the tool. A lawyer who knows exactly what a clause should say can get one drafted. They just can't do it with less understanding than before. They can do it with different understanding.\n\nCorn: Which quietly kills the fantasy. The tool doesn't make the domain knowledge unnecessary. It makes the syntax the easy part.\n\nHerman: And syntax was never the hard part for anyone who'd been doing it long enough to be useful.\n\nCorn: Daniel asked specifically about three groups: established developers, people learning, and complete newcomers vibe-coding. How cleanly can we size those?\n\nHerman: Not cleanly at all, and I want to be straight about that. Anthropic's classifier measures expertise within a task. It isn't a proxy for career stage. There's no reliable breakdown of learners versus newcomers as distinct cohorts, so nobody can give you a percentage split with a straight face.\n\nCorn: What the data does show is that it's widening who attempts software without making anyone equally effective.\n\nHerman: That's the fairest single sentence on the whole subject. Domain experts in law, finance and management succeed nearly as often as engineers. Novices succeed at half the rate and quit at three to four times the rate.\n\nCorn: So the door is open. Not everyone who walks through it comes out the other side.\n\nHerman: Now. What are people building?\n\nCorn: That's the question I suspect has the least satisfying answer.\n\nHerman: It does. Let me give you the measurable part first. The share of GitHub projects with coding-agent activity more than doubled since late last year. A study covering a hundred and twenty-eight thousand public repositories found agent activity in an estimated sixteen to twenty-three percent of projects as of the end of October, and adoption was more than twice as high among projects created after that period.\n\nCorn: New projects are the ones going agent-first. That makes sense. You don't have to retrofit anything.\n\nHerman: And the headline number: four percent of all public commits on GitHub worldwide were authored by Claude Code in February. Double the prior month.\n\nCorn: Four percent of every commit on GitHub. That's not a niche anymore.\n\nHerman: It's the largest single share any tool has had. Their own engineering is the extreme case. The majority of their internal code is now written by Claude Code, and this year is the first time internal pull requests inflected upward because of it.\n\nCorn: Their engineers got more productive and the graph bent upward, which is not how productivity headlines usually go.\n\nHerman: Now the unsatisfying part. Anthropic says explicitly they can't observe whether code written in a session is used or discarded. And they exclude non-interactive usage, the headless runs, which they describe as a substantial share of activity.\n\nCorn: So the measurement stops at the edge of the interesting question. Nobody knows how much of it ships.\n\nHerman: You can see the estimated economic value of the average session rising, twenty-seven percent between October and April, building up forty-three, operating thirty-four, fixing thirty-two. But that's value measured against freelance job postings. It's an estimate of worth, not a count of deployments.\n\nCorn: So what are people building? Conventional software.\n\nHerman: Overwhelmingly. Internal tools, automations, prototypes, the plumbing of businesses, and a growing share of people operating software they already had rather than writing anything new. That's the plain answer. The romantic answer would be entirely new categories, and there are some, but the mass is ordinary software getting made faster.\n\nCorn: Which is the least interesting answer and the most consequential. If ordinary software gets made five times faster, that's a bigger deal than a novelty app.\n\nHerman: That's the whole economic argument in one line.\n\nCorn: Why did this one take off, though? There were coding assistants before it, and Copilot had a two-year head start and Microsoft's distribution.\n\nHerman: Two reasons, and they're not the same reason. The first is architecture. Claude Code reads a codebase, plans a sequence of actions, executes them with real development tools, evaluates the result, and adjusts. That loop is the product. An autocomplete suggests a line. This thing holds a task.\n\nCorn: So it's the difference between someone finishing your sentence and someone finishing your job.\n\nHerman: That's it. And the usage patterns show it. In Anthropic's earlier index work, seventy-nine percent of Claude Code conversations were classified as automation rather than augmentation. On Claude dot ai, it's the inverse. People aren't having a conversation with it. They're assigning work.\n\nCorn: That's a strange thing to measure. What separates the two?\n\nHerman: Whether the human stays in the loop for each step. Augmentation is you and the model trading turns. Automation is you describe the outcome and walk away.\n\nCorn: A machine that does the work while nobody is watching. That's very close to my personal ideal and I'm choosing not to examine why.\n\nHerman: I'd examine it, honestly. The second reason it took off is model quality. In the Pragmatic Engineer survey, Opus 4.5 and Sonnet 4.5 were mentioned more often for coding than every other model combined.\n\nCorn: So the harness matters and the engine matters.\n\nHerman: The harness is why people stay and the engine is why they arrive. You don't get eighty percent conversion from regular use to primary tool on architecture alone.\n\nCorn: And the architecture is a solved-ish problem now. Everybody's building the loop.\n\nHerman: Which means the moat is the model and the moat is thinner than the revenue line suggests. That's my suspicion, anyway. Cursor had the loop early and lost ground when someone else's model got better.\n\nCorn: Now the bit that's been brewing under all of this. The backlash.\n\nHerman: It's real and it's measurable. Uber reportedly burned its entire twenty twenty-six AI budget on Claude Code in four months. Microsoft started canceling Claude Code licenses in May. Alibaba moved to ban it in July. Accenture capped Claude spending at twenty-nine dollars a month this month.\n\nCorn: Twenty-nine dollars a month is not a cap, that's an insult with a budget code.\n\nHerman: It's procurement saying \"not this year\" in the most polite available language. And it's the knock-on effect of the adoption curve nobody plans for. The tool is priced to reward usage, and usage, once it becomes the primary tool, scales with the size of the engineering org.\n\nCorn: The fastest-growing users become the fastest-growing bill, and finance notices.\n\nHerman: Every company that adopts it enthusiastically for six months discovers the invoice in month seven. That doesn't kill the tool. It changes who uses it and how carefully.\n\nCorn: It changes the shape of the curve. Not a collapse, a negotiation.\n\nHerman: Which brings the whole thing to the question Daniel actually asked underneath the other questions. Has it widened the population of people who can build software, or is the impact still concentrated among existing developers?\n\nCorn: The evidence says both, and the both is the interesting part.\n\nHerman: It widened who attempts. Non-software occupations are the fastest-growing groups, and a finance person succeeds nearly as often as an engineer when the task is code. But it didn't make anyone equally effective. Novices succeed at half the rate and abandon at three to four times the rate.\n\nCorn: It's a tool that opens the door and then charges admission in the form of understanding.\n\nHilbert: I agree with you.\n\nHerman: Good. That's settled.\n\nHilbert: I have a repository. Private, it's a ledger thing for the building's service charges, I've maintained it since before any of this. Four hundred and sixteen commits. I ran the agent over it for a weekend in July and it made eleven thousand two hundred commits. In two days. To a repository with four hundred commits in it.\n\nCorn: Eleven thousand commits to a service charge ledger.\n\nHilbert: The number I keep coming back to is the ratio. Nine hundred to one. That's the real adoption statistic, nobody's publishing that one. I've been at this desk all morning with it.\n\nHerman: How do you even generate eleven thousand commits in two days? That's a commit every fifteen seconds.\n\nHilbert: It committed every time it changed its mind. Some of them said \"rename variable.\" It would rename it back an hour later and commit that too. It didn't necessarily get anywhere, it just kept a record of the walking.\n\nCorn: The commit log is the transcript of the thinking, and the thinking was mostly walking.\n\nHilbert: The history was two hundred and forty megabytes. I couldn't open it. Herman, you'd have looked at the log and diagnosed it in a minute, this is your sort of thing.\n\nHerman: I would have looked at the log and gone to bed.\n\nHilbert: The repository is gone now. I deleted it in August. Not because it was broken. Because I couldn't find the commit where it stopped being mine.\n\nCorn: Hm.\n\nHilbert: The service charges got done, though. Two hundred and eleven apartments, all reconciled, and the man in flat nine stopped writing me letters. That's the part I'd keep.\n\nCorn: So, what's still open?\n\nHerman: Three things. The user count, which nobody has published. The production readiness question, which nobody can currently measure. And whether the cost backlash is the leading edge of a plateau or just a correction inside a curve that's still steep.\n\nCorn: The one that would change the most is the next twelve months of user mix. If the fastest-growing groups really are management, sales and legal, then we're about to find out whether this is a developer tool that spread or a general-purpose software-creation tool that happened to start with developers.\n\nHerman: They're different products with different trajectories. A developer tool hits a ceiling at the population of developers. A general-purpose tool doesn't.\n\nCorn: The ceiling nobody can see is still the one that matters. Anthropic will publish revenue to the last decimal. They won't publish the number of people holding the thing.\n\nHerman: Hilbert, thank you for producing.\n\nCorn: If you enjoyed this one, try episode twelve, The AI Breakthrough; episode twenty-four sixty-one, How Claude Code's Conversation Compaction Actually Works; and episode fourteen sixty-four, When AI Agents Author four% of All Code. This has been My Weird Prompts. If you've got a minute, leave us a review, it helps people find the show.\n\nHerman: We'll be back soon.\n\nCorn: See you then.","showNotes":"Claude Code's rise is real, but the numbers people cite are a mix of disclosures, estimates, and noise. Anthropic has never published an absolute user count — only growth multiples like \"weekly active users doubled since January\" — while revenue figures are steep and verifiable: $500M run-rate in September 2025, $1B by December, $2.5B by February, and $16.8B reported to investors this month. Third-party estimates and a widely-shared \"2 million weekly users\" figure trace back to no Anthropic statement at all. Surveys disagree sharply depending on methodology: JetBrains found 39% of professional developers using Claude Code at work (47% in the US), while Stack Overflow found 65.5% had used it in the past year — different questions, different populations. Both agree Copilot's lead eroded. Adoption inverts by company size: 75% at firms of 1-50 people, but Copilot leads at 56% in companies over 10,000, where procurement favors the incumbent vendor. An Anthropic study of ~235,000 people found engineers lead adoption but management, sales, and legal are the fastest-growing groups. Sessions are drifting from writing code toward operating systems. And the finding that cuts against the vibe-coding pitch: novices reach verified success 15% of the time versus 28-33% for intermediate and above, abandoning troubled sessions at 19% versus 5-7%. Yet across occupations that produce code, success rates land within seven points of software engineers — management highest of all. Domain expertise, not coding background, may be what agents reward.","promptSanitized":"The Explosive Rise of Claude Code: Who's Using It, and What Are They Building?\n\nOver the past six to twelve months, Claude Code appears to have gone from a relatively specialized AI coding tool to an enormous phenomenon. It seems to have captured the attention of professional software developers, hobbyists, entrepreneurs, and people who have never written a line of code before.\n\nI'd like to explore just how dramatic this growth has actually been, who's driving it, and what people are doing with the technology.\n\nFirst, how big has Claude Code become? What do we know about its adoption trajectory, estimated active users, revenue, and growth over the past year? What reliable figures has Anthropic disclosed, and what can we reasonably estimate from other sources? How does its adoption compare with GitHub Copilot, Cursor, and other AI coding tools?\n\nSecond, who is actually using Claude Code? I'd be particularly interested in distinguishing three groups: established software developers incorporating it into professional workflows; people using it to learn software development; and complete newcomers taking a predominantly vibe-coding approach. Do we have any evidence about the relative sizes of these groups, or how their usage patterns differ?\n\nThird, what are people building? Are most users working on conventional software projects, websites, internal tools, automations, and prototypes? Or are we seeing entirely new categories of projects that previously wouldn't have been practical for individuals to attempt? What do we know about the scale, complexity, and production readiness of these projects?\n\nFinally, why has Claude Code become such a phenomenon? What distinguishes it from earlier coding assistants, and how much of its success comes from improvements in the underlying models versus its agentic architecture and user experience? Has it genuinely expanded the population of people capable of building software, or is its impact still concentrated among existing developers?\n\nI'd like to distinguish measurable adoption from social-media hype throughout, while exploring whether Claude Code represents a fundamental shift in who creates software and how it gets built.","transcriptUrl":"https://episodes.myweirdprompts.com/transcripts/claude-code-adoption-data-37156.md","url":"https://myweirdprompts.com/episode/claude-code-adoption-data-37156/","_links":{"self":"https://myweirdprompts.com/api/episodes/claude-code-adoption-data-37156.json","html":"https://myweirdprompts.com/episode/claude-code-adoption-data-37156/","audio":"https://episodes.myweirdprompts.com/audio/claude-code-adoption-data-37156.m4a"}}