Which AI coding tool makes the most money per user?
In our AI code assistant market deck, you will find everything you need to understand the market
SUMMARY
Claude Code is the best-supported answer to the main question: which AI coding tool makes the most money per user? Among major products where both revenue and active usage can be estimated, Anthropic appears to capture the most annualized revenue per active developer.
The headline gap is large enough to survive imperfect public data. Claude Code works out to roughly $1,250 of annualized revenue per reported weekly active user using the best available figures, while even a substantially larger user base would still leave it in the high hundreds of dollars per active user.
Cursor is probably the closest measurable rival. Its roughly $4 billion annualized revenue run rate suggests unusually strong monetization, but its freshest revenue figure has outrun the public monthly-active-user data, so a precise current ARPU would be fake precision.
The market is splitting into two economic models. Claude Code, Cursor and Devin increasingly monetize professional engineering work, while Lovable, Replit and Base44 deliberately serve much broader populations that include occasional builders, students, founders and non-technical users.
That audience difference matters more than raw scale. A product with a few million serious developers can make far more per user than one with tens of millions of registrations if those developers are using agents for hours of paid work every week.
GitHub Copilot shows the opposite strength: enormous distribution with monetization still catching up. Its 50 million users dwarf the specialist tools, but Microsoft only recently began showing how much usage-based billing can lift revenue from the heavy users already inside that funnel.
Devin is the biggest unresolved case. Cognition has pushed annualized revenue above $900 million and sells expensive autonomous engineering capacity to enterprises, but without a credible current user or seat count there is no defensible revenue-per-user figure.
Usage-based billing is raising the ceiling across the category. AI coding companies no longer have to grow only by adding seats; they can also earn more when each developer delegates more work and consumes more model capacity.
High revenue per user does not automatically mean the best business. These products can carry heavy inference costs, so the durable winner will eventually need both strong monetization per developer and enough gross margin to keep a meaningful share of that revenue.
The most useful takeaway is that user count has become a weak standalone scoreboard for AI coding. The more important question is how much valuable engineering work each user delegates, because that is increasingly what determines how much revenue a product can capture.
This market map, featured in our AI code assistant market deck, highlights top companies and startups in the AI code assistant market
Why does revenue per user suddenly matter so much for AI coding tools?
AI coding tools have become billion-dollar businesses so quickly that user count alone now tells us surprisingly little about who is really winning.
Cursor's annualized revenue reportedly reached about $4 billion by June, up from $2 billion only a few months earlier. Anthropic had already disclosed more than $2.5 billion of run-rate revenue for Claude Code earlier in the year. Cognition, the company behind Devin, went from $492 million in annualized revenue in late May to more than $900 million by the end of August. Lovable passed $500 million. Base44 recently crossed $200 million.
At the same time, these products serve wildly different populations. GitHub Copilot now reaches 50 million users. Replit says more than 50 million people use its platform. Base44 has more than 10 million users. Cursor had roughly seven million monthly active users at the last widely reported estimate, while Claude Code's weekly audience has been reported at around two million.
That creates a strange market. A tool can have fewer users and still make several times more money from each one.
Revenue per user is useful precisely because AI coding is splitting into two models. Products such as Lovable and Replit want almost anybody to build software. Claude Code, Cursor and Devin are increasingly selling expensive AI work to professional developers and large companies.
The second group has a much smaller audience, but every serious user can be worth hundreds or even thousands of dollars a year.
What does “money per user” actually mean for Claude Code, Cursor and GitHub Copilot?
The best measure is annualized product revenue divided by active human users, but no major AI coding company publishes that exact pair of numbers consistently.
Registered users are especially dangerous here. Someone who tried Replit at school three years ago and never returned still counts toward a cumulative user figure. A developer using Claude Code five days a week is economically very different.
Paying users create another problem. GitHub disclosed 4.7 million paid Copilot subscribers earlier this year, while Anthropic and Cursor have not released directly comparable paying-user counts. Enterprise contracts also make the idea of a “user” fuzzy because one company can buy thousands of seats and then consume additional AI usage on top.
We therefore use active users whenever a credible figure exists. Where it does not, we use the closest public denominator and say clearly what it measures.
This avoids one of the easiest mistakes in the market: dividing fresh revenue by an old cumulative registration number and calling the result ARPU.
As this chart shows, and as featured in our AI code assistant market deck, search interest in AI code assistants has increased significantly
Which AI coding tools actually give us enough data to compare?
Claude Code and Cursor give us the best shot at estimating revenue per active developer, while GitHub Copilot, Lovable, Base44 and Replit are useful mainly for showing how different the economics can become.
Anthropic disclosed more than $2.5 billion in Claude Code run-rate revenue and said weekly active users had doubled early in the year. A later figure of roughly two million weekly users has circulated widely, although Anthropic has not formally published that exact number. We should therefore treat the resulting ARPU as an order-of-magnitude estimate rather than an audited metric.
Cursor's latest revenue figure is fresher. Forbes reported that annualized revenue passed $4 billion in June, after reaching $3 billion in late April and $2 billion around February. The best widely available monthly-user estimate, roughly seven million, comes from earlier in the year. Cursor also had more than one million daily active users by the end of March, according to a JFrog announcement tied to its marketplace launch.
GitHub gives us stronger user data but weaker revenue data. Microsoft now says GitHub Copilot has 50 million users. Earlier in the year it disclosed 4.7 million paid subscribers. Microsoft still does not report standalone Copilot revenue.
Cognition is the opposite problem. Revenue is now reported above $900 million annualized, but Cognition does not disclose enough current Devin users or seats for us to calculate a clean number.
| AI coding tool | Latest useful revenue figure | Best public user figure | How usable is it? |
|---|---|---|---|
| Claude Code | $2.5B+ run rate | ~2M weekly users reported | Good for an approximate range |
| Cursor | ~$4B annualized | ~7M MAU earlier in year | Useful, but dates do not match |
| GitHub Copilot | Not disclosed separately | 50M users; 4.7M paid earlier | Good denominator, missing revenue |
| Cognition / Devin | $900M+ annualized | No clean current figure | Cannot calculate |
| Lovable | $500M annualized | Millions of users | Broad-user comparison only |
| Base44 | $200M+ ARR | 10M+ users | Broad-user comparison |
| Replit | ~$525M estimated | 50M+ users | Broad-user comparison |
Is Claude Code making the most money per active user right now?
Claude Code is still the best-supported answer today, with an indicative annualized revenue figure around $1,250 per weekly active user.
The calculation comes from Anthropic's disclosed Claude Code run rate of more than $2.5 billion and the later reported figure of roughly two million weekly active users. That works out to about $104 per active user per month.
We would not pretend $1,250 is Claude Code's exact ARPU. The two data points come from different moments, the two-million-user figure is not an Anthropic disclosure, and Claude Code has kept growing since both numbers appeared.
What makes the estimate convincing is the scale of the gap. Even if Claude Code had three million weekly users instead of two million, $2.5 billion of annualized revenue would still equal more than $830 per weekly user. At four million, it would still be $625.
Anthropic's own disclosures point the same way. Claude Code's run rate more than doubled early in the year, while weekly active users doubled. Business subscriptions quadrupled over the same stretch, and enterprise customers already accounted for more than half of Claude Code revenue.
Revenue was therefore growing much faster than the most directly comparable usage metric.
Claude Code also sits unusually deep inside the developer's workday. Reporting around the product has put average usage near 20 hours per week. When developers use a coding agent for that much of their working time, charging $100 or more per month stops looking unusual.
If you want more recent data on this point, please see our latest AI code assistant market report.
This chart, featured in our AI code assistant market deck, illustrates yearly VC funding for AI code assistant startups
How much money does Cursor make per user now?
Cursor now appears to generate somewhere in the high hundreds of dollars per active user annually, although its latest $4 billion revenue figure has outrun the public user data we need for a precise calculation.
If we mechanically divide $4 billion by the earlier estimate of seven million monthly active users, we get about $571 per user per year, or roughly $48 per month.
That probably overstates today's true figure because Cursor's user population has grown since the seven-million estimate. India alone recently reached around three million Cursor developers after more than tripling over a year. The country is now Cursor's third-largest market.
Still, even a much larger denominator leaves impressive economics. At 10 million monthly users, $4 billion would equal $400 per user annually. At 15 million, it would still be roughly $267.
Cursor's pricing makes those numbers believable. Pro costs $20 a month, heavier individual tiers run much higher, and team plans can cost $40 or $120 per seat before extra consumption. Cursor has also said daily Agent users commonly consume around $60 to $100 of model usage per month, with some power users going above $200.
The bigger change has happened inside companies. Corporate customers accounted for roughly 60% of revenue when Cursor was around the $2 billion run-rate mark, and more than 50,000 enterprises now use the product. Cursor says 64% of the Fortune 500 has adopted it.
The latest revenue trajectory makes the shift even clearer. Cursor went from roughly $1 billion annualized late last year to $2 billion, then $3 billion and finally $4 billion within months. User growth alone is unlikely to explain that jump. Bigger enterprise contracts and heavier AI consumption are doing a lot of the work.
Why does Claude Code make more money from each developer than Cursor?
Claude Code can charge more per developer because Anthropic has pushed heavy users toward expensive subscriptions and consumption-based usage while targeting people who use the agent for serious engineering work.
Claude starts at $20 a month for Pro, but Max plans cost $100 or $200. Premium team seats can also reach around $100 or more per user each month. Enterprise customers increasingly pay for a seat plus the model usage generated by the developer.
That gives Anthropic room to collect more money when someone asks Claude Code to do more work.
A developer using an old-style coding assistant might accept completions, ask a few questions and generate a function. A Claude Code user can hand over a repository, ask the agent to inspect it, change several files, run tests, debug failures and continue working for a long stretch.
More work means more inference, but it can also mean more economic value for the customer. Spending $150 in a month looks expensive next to a $20 SaaS seat. It looks far less expensive if the agent saves several hours of engineering time every week.
Claude Code's customer mix confirms that companies are willing to pay this way. More than half of the product's revenue already comes from enterprise users, and Anthropic said business subscriptions quadrupled during the same period in which weekly users doubled.
Cursor is moving toward similar economics, but Claude Code appears further along in turning intense agent usage into revenue per developer.
If you want more recent data on this point, please see our latest AI code assistant market report.
This chart, featured in our AI code assistant market deck, breaks down Anyshpere’s playbook in AI code assistants
Is GitHub Copilot making surprisingly little money from its 50 million users?
GitHub Copilot almost certainly makes far less revenue per total user than Claude Code, but Microsoft's latest numbers show that Copilot monetization is speeding up.
Microsoft said GitHub Copilot reached 50 million users in its latest annual earnings call. Earlier in the year, GitHub had 4.7 million paid Copilot subscribers, up 75% year over year.
The gap between those two figures tells us plenty on its own. Copilot reaches more than ten times as many people as its earlier paid-subscriber count.
Microsoft still does not disclose standalone GitHub Copilot revenue, so an exact ARPU would be guesswork. The useful new information is what happened after GitHub changed its billing model.
GitHub introduced usage-based billing in June. In Microsoft's latest quarter, Copilot revenue then accelerated by more than 60% quarter over quarter. Microsoft also said the stronger-than-expected consumption helped its Intelligent Cloud results.
That is a much more important development than another user milestone. Copilot already had huge distribution. The missing piece was getting more revenue from each serious user.
GitHub Business and Enterprise customers now pay for seats while heavy model use can create extra consumption revenue. Microsoft is starting to monetize Copilot more like Claude Code and Cursor.
Claude Code still looks much more revenue-dense today. GitHub has something Anthropic does not: tens of millions of existing users who can gradually be moved into paid and usage-based plans.
Could Devin actually make more money per user than Claude Code?
Devin could already be the highest-monetizing AI coding product among large enterprise customers, but Cognition still does not publish enough user data for us to prove it.
The revenue curve has become hard to ignore. Cognition said annualized revenue reached $492 million in late May. By the end of August, The Information and Bloomberg were reporting more than $900 million.
That means annualized revenue almost doubled in roughly three months.
Cognition had already said enterprise usage of Devin was growing around 50% month over month for six consecutive months before the $492 million milestone. Customers include Goldman Sachs, Mercedes-Benz, Santander, BNY, Dell and other large organizations.
Devin's economics also allow much larger bills than a normal coding subscription. Cognition sells autonomous engineering work through metered agent usage, and enterprise contracts can reach very large amounts. A Cognition Platform listing on AWS Marketplace has shown a $1 million 12-month contract dimension before additional Agent Compute Units.
We should not extrapolate that price across Cognition's entire customer base. It does show how high spending can go when companies buy autonomous engineering capacity instead of a $20 developer tool.
The missing denominator blocks the ranking. Cognition could have 100,000 intense Devin users, 500,000, several million, or a very different mix once Windsurf is included. Each scenario produces a completely different revenue-per-user figure.
For now, Claude Code wins on the numbers we can actually calculate. Devin is the obvious company that could upset that answer once Cognition gives us better user data.
If you want more recent data on this point, please see our latest AI code assistant market report.
This chart, featured in our AI code assistant market deck, illustrates yearly funding for AI code assistant startups
Why do Lovable, Base44 and Replit make much less money per user?
Lovable, Base44 and Replit generate much less revenue per broad user because they deliberately serve millions of occasional builders and non-developers, not only highly paid engineers using an agent every day.
Lovable passed $500 million in annualized revenue in June and was already producing around one million new projects every week. The company now says more than 60 million projects have been built on the platform.
Base44 gives us an even cleaner recent comparison. It passed $200 million ARR in August, five months after reaching $100 million, while exceeding 10 million users. That puts broad-user revenue at less than $20 a year.
Replit says more than 50 million people use its platform. Sacra has estimated its annualized revenue at roughly $525 million. Against that enormous cumulative audience, the result is only around $10 per user annually.
Those figures sound tiny beside Claude Code, but the comparison says more about who uses the products than about product quality.
Lovable's own research has repeatedly shown that many customers are founders, operators and non-technical builders. Replit has expanded beyond programmers into students, entrepreneurs and employees building internal tools. Base44 sells the idea that almost anybody can turn an idea into working software.
A huge free funnel naturally drags down revenue per broad user.
Claude Code begins much closer to the money. Someone installing a terminal-based coding agent and using it throughout the workweek is far more likely to have a professional reason to pay.
| Product | Latest annualized revenue | Broad user figure | Approx. revenue per broad user |
|---|---|---|---|
| Lovable | $500M | Millions; exact current denominator unclear | Below professional coding-agent levels |
| Base44 | $200M+ | 10M+ | < $20/year |
| Replit | ~$525M estimated | 50M+ | ~ $10/year |
Is usage-based billing pushing AI coding revenue per user higher?
Usage-based billing is now one of the biggest forces pushing AI coding revenue per user upward because heavy users can keep spending after the initial subscription is sold.
GitHub's latest results are the clearest fresh example. Microsoft introduced usage-based Copilot billing in June and then reported more than 60% sequential revenue acceleration for the product. The company explicitly credited stronger Copilot consumption for part of its quarterly performance.
Cursor has gone down the same road. Teams buy seats with included model usage, then pay more when developers consume additional capacity. Cursor's own estimates put frequent Agent users around $60 to $100 of monthly usage, with some power users exceeding $200.
Anthropic already sells Claude Code into enterprises through a combination of access and model consumption. A developer who runs many long agent sessions can therefore be worth several times more than one who only opens Claude Code occasionally.
Cognition goes furthest. Devin customers are effectively buying quantities of autonomous engineering work.
Traditional SaaS usually tries to add more seats. AI coding companies can add seats, increase usage per seat and move individual developers into more expensive models at the same time.
The revenue opportunity per developer can therefore keep climbing even if user growth eventually slows.
This chart, featured in our AI code assistant market deck, compares the main business model options for AI developer tools platforms
Does the ranking change if we count only paying users?
Revenue per paying user would produce a different ranking, but the public data is too uneven to tell us who wins.
GitHub gives us one of the cleanest figures: 4.7 million paid Copilot subscribers earlier in the year. Unfortunately, Microsoft does not give us standalone Copilot revenue.
Lovable gave us a rare useful snapshot in 2025. At the time, the company had more than 180,000 paying subscribers and roughly $100 million ARR. That worked out to around $556 per paying subscriber annually, or about $46 a month.
Yet Lovable simultaneously had millions of active users, so revenue per active user was far lower.
The same denominator problem becomes even larger for Claude Code. A user may pay through Claude Pro, Max, Team, Enterprise or consumption attached to a company account. Anthropic does not disclose how many individual paying Claude Code users sit underneath the product's $2.5 billion-plus run rate.
Cursor combines individual plans, team seats, enterprise contracts and extra model usage as well.
A clean paid-user leaderboard would therefore require private company data. We can make a solid active-user judgment, but pretending we know exact paid-user ARPU across these companies would add fake precision.
If you want more recent data on this point, please see our latest AI code assistant market report.
Does making more revenue per user mean Claude Code has the best business?
Claude Code's high revenue per user shows strong willingness to pay, but AI coding agents also carry much heavier computing costs than ordinary software.
Cursor demonstrated the problem early. Reporting around the company indicated that it had operated with negative gross margins because some users consumed more third-party model inference than their subscriptions covered.
Cursor has spent much of this year fixing that. Proprietary models such as Composer, cheaper routing options and more disciplined usage pricing helped the business move toward positive gross margins. Enterprise sales reportedly became gross-margin positive before some individual accounts did.
Microsoft gave us another useful clue in its latest results. GitHub Copilot consumption rose sharply after usage-based billing was introduced, but Microsoft also said growing AI usage weighed on gross margins. The company added that Copilot margins improved through the quarter after the billing change.
Anthropic faces the same basic issue. Claude Code users can run expensive models for hours, so part of that impressive revenue per developer goes straight back into inference.
The best AI coding business will eventually depend on two numbers: revenue per user and the compute cost required to generate it.
Claude Code looks excellent on the first one. We do not have enough standalone product data to compare the second one properly.
This chart, featured in our AI code assistant market deck, illustrates how market revenue is distributed across customer segments in the AI code assistant market
Which AI coding user numbers are most likely to mislead us?
Cumulative users are the most misleading number in this market, followed closely by revenue and user figures taken from different periods.
Replit's 50 million-plus users span years of product history. That population includes current professionals, students, free users and people who may no longer use Replit at all.
GitHub's 50 million Copilot users describe a much broader group than the 4.7 million paid subscribers Microsoft disclosed earlier. Using either denominator without explaining the difference can change the apparent ARPU by more than tenfold.
Cursor creates a different problem. Its $4 billion annualized revenue figure is fresh, while the commonly cited seven-million monthly-active-user estimate is older. Dividing one by the other probably makes Cursor's current ARPU look too high.
Claude Code has a similar timing issue in reverse. Anthropic formally disclosed more than $2.5 billion of run-rate revenue earlier in the year, while the later figure of roughly two million weekly users has not been confirmed directly by the company.
As seen above, the gaps between these products are sometimes large enough that the conclusion survives the messy data. Claude Code at roughly $800 to $1,250 per active user would still sit in a very different range from Base44 or Replit at tens of dollars per broad user.
What we should avoid is pretending that a $570 estimate is meaningfully more accurate than $500 when the underlying denominator itself is moving fast.
So which AI coding tool makes the most money per user today?
Claude Code is the best-supported winner today: among major AI coding products where we can estimate both revenue and active usage, Anthropic appears to make the most money per active developer.
Using the best public figures available, Claude Code lands around $1,250 of annualized revenue per reported weekly active user. Even after allowing for a substantially larger user base, the figure remains in the high hundreds of dollars.
Cursor is probably next among the tools we can measure. Its latest revenue run rate has climbed to roughly $4 billion, and a reasonable range for today's user base would still leave revenue per monthly active developer somewhere in the hundreds of dollars.
GitHub Copilot reaches far more people, with 50 million users, but Microsoft is only now showing how much usage billing can raise monetization. The latest quarter was striking: Copilot revenue accelerated by more than 60% sequentially after the pricing change.
Lovable, Base44 and Replit sit much lower on broad-user revenue because their funnels include millions of free and occasional builders. That is part of their strategy.
Devin remains the big unresolved case. Cognition has gone from $492 million to more than $900 million in annualized revenue within a few months, and its enterprise contracts can be enormous. Without a reliable active-user count, we cannot tell whether Devin makes $300, $1,000 or several thousand dollars per serious user.
So if someone asks which AI coding tool currently makes the most money per user, our answer is Claude Code.
The more interesting race now is between Claude Code, Cursor and Devin. All three are learning that a professional developer can be worth far more than a $20 monthly software seat once the product starts doing hours of engineering work on that person's behalf. The winner may eventually be the company that gets each developer to delegate the most valuable work, rather than the company that simply collects the most users.
If you want more recent data on this point, please see our latest AI code assistant market report.
This chart, featured in our AI code assistant market deck, shows how AI coding assistant technology has evolved over time
OUR METHODOLOGY
We approached this question as a measurement problem rather than a simple leaderboard. Public numbers for AI coding tools are unusually difficult to compare because companies report revenue at different moments, describe users in different ways, and increasingly combine subscriptions, enterprise contracts and usage-based billing.
For each company, we collected the freshest first-party disclosures and high-quality reported figures we could find, then assessed them together rather than allowing one headline number to decide the ranking. We prioritized active usage over cumulative registrations, separated paid users from broader audiences, and treated weekly, monthly and total-user figures as different measures rather than interchangeable versions of “users.”
We only calculated revenue per user where the public data was strong enough to support it. Where a company had compelling revenue data but no usable denominator, as with Devin, we kept it in the analysis without manufacturing a number. Where revenue and usage figures came from different periods, we tested several plausible user-base sizes to see whether the conclusion survived instead of attaching false precision to one calculation.
Pricing, enterprise adoption, usage intensity and consumption-based billing were used as explanatory evidence. They help distinguish products monetizing occasional access from products monetizing hours of delegated engineering work, but we did not use them as substitutes for actual revenue and user data.
We also kept monetization separate from profitability. High revenue per developer tells us how much customers are willing to spend; it does not tell us how much of that revenue remains after model inference and other compute costs. The ranking therefore answers a deliberately narrow question: which major AI coding product appears to capture the most revenue per active user based on the strongest comparable evidence available.
Key sources include Anthropic on Claude Code's run-rate revenue, user growth and enterprise mix, Anthropic's Claude Code usage research, Anthropic pricing, Microsoft FY2026 Q2 on paid GitHub Copilot subscribers, Microsoft FY2026 Q4 on Copilot users, consumption and revenue acceleration, GitHub on the move to usage-based Copilot billing, and GitHub's Copilot plan documentation.
For Cursor, we relied on Forbes on Cursor's roughly $4 billion annualized revenue, JFrog on more than one million daily active Cursor users, Cursor's enterprise adoption figures, Cursor's pricing and usage economics, Cursor's India developer figures, and The Information on Cursor's gross-margin pressure and improvement.
For Cognition, Lovable, Base44 and Replit, key sources include The Information on Cognition's annualized revenue, AWS Marketplace on Cognition Platform contract structure, TechCrunch on Lovable's $500 million annualized revenue, Lovable on projects built, Wix on Base44's ARR progression, and Replit on its global user base and audience breadth.
Our objective was not to force every company into a cosmetically complete table. It was to aggregate the strongest recent signals, test the comparisons that could materially change the answer, and make the final judgment only where the evidence was strong enough to support it.
In our AI code assistant market deck, we identify pain points entrepreneurs should prioritize
Related blog posts
- Which AI coding startups are generating the most revenue today?
- Who is winning the AI coding market: Cursor, Claude Code, or Copilot?
