Which AI tools have the strongest retention now?

Are AI tools actually good at keeping paying users?

AI tools are still surprisingly bad at retention overall, despite being unusually good at getting people to pay in the first place.

RevenueCat's latest analysis of more than 115,000 subscription apps makes the gap hard to ignore. AI-powered apps generated 41% more realized value per payer over the first year than non-AI apps, at $30.16 versus $21.37. Yet AI retention was worse at every main subscription length.

After twelve months, only 6.1% of monthly AI subscriptions were still retained, compared with 9.5% for non-AI apps. Annual plans showed an even larger absolute gap: 21.1% for AI versus 30.7% for non-AI.

AI apps also lean heavily on monthly subscriptions. RevenueCat found that 59.8% of AI subscriptions were monthly, versus 26.2% for non-AI products. That gives users a frequent opportunity to ask the obvious question: “Am I still using this enough to pay again?”

So the bar for calling an AI product genuinely sticky should be high. Huge signup numbers, fast ARR growth and strong trial conversion tell us very little about whether customers are still around six or twelve months later.

Metric	AI apps	Non-AI apps
12-month monthly-plan retention	6.1%	9.5%
12-month annual-plan retention	21.1%	30.7%
Year-one value per payer	$30.16	$21.37
Median refund rate	4.2%	3.5%

What should we actually count as strong AI retention?

Strong AI retention means people repeatedly choose the same product after they have had enough time, money and alternatives to leave it.

That sounds simple until we compare products.

For ChatGPT or Claude Pro, paid subscription survival is useful because an individual can cancel easily. For Cursor, net revenue retention tells us whether an existing customer cohort is spending more or less over time. For enterprise AI, purchase persistence matters, but so does actual usage: a company can keep thousands of unused seats because cancelling them is somebody else's problem.

Coding agents add another layer. A developer opening Claude Code every working day shows much stronger behavioral retention than someone who keeps a $20 subscription but visits twice a month.

We therefore need several tests. Does the customer keep paying? Does usage continue? Does spending expand? Does the product survive when the customer also tries a competitor? And has the AI become attached to something the user repeatedly needs to do?

There is no credible single-number leaderboard. The interesting part is seeing which products stay near the top across several of those tests.

Does ChatGPT have the strongest consumer AI retention today?

ChatGPT currently has the strongest demonstrated mass-market consumer retention among standalone AI products.

The cleanest public cohort we found comes from YipitData's U.S. e-receipt panel. It followed 40,857 people who were paying for ChatGPT at the end of 2025 and were not simultaneously paying for Claude. Six months later, 80.4% still showed an active ChatGPT subscription.

The methodology can miss some direct cancellations, so 80.4% should not be read as a perfect audited churn number. What makes it useful is the size of the cohort and the persistence it shows.

ChatGPT also has the scale to make that number unusually meaningful. OpenAI says ChatGPT now has more than one billion weekly active users. Earlier this year, it had already reported more than 50 million consumer subscribers. A niche professional product can retain a highly selected group of users; ChatGPT has to keep people ranging from programmers and students to casual consumers.

Andreessen Horowitz reached a similar conclusion from another angle. Its latest consumer-AI analysis said ChatGPT and Gemini were producing best-in-class consumer paid retention, while ChatGPT still generated substantially more sessions per user: around 1.3 times Gemini on the web and 2.2 times as many on mobile.

For broad consumer AI, ChatGPT is the clearest retention leader we can defend today.

Is Claude actually stickier than ChatGPT for serious AI users?

Claude looks exceptionally sticky among people who have actually paid for both Claude and ChatGPT, and that is probably the strongest evidence in Anthropic's favor.

YipitData followed 2,439 U.S. consumers who were paying for ChatGPT and Claude at the same starting point. Six months later, roughly 92% still showed an active paid Claude signal, either alone or alongside ChatGPT.

The head-to-head behavior is even more interesting. By the end of the measurement period, 24.7% of the original dual subscribers had kept Claude while dropping ChatGPT. Only 5.0% had kept ChatGPT while dropping Claude.

Claude was therefore about five times more likely to be the sole survivor when these dual subscribers consolidated.

We should be precise about whom this describes. Someone willing to pay for two frontier AI assistants is already an unusually heavy AI user. We cannot take that 92% and apply it to every Claude subscriber.

But this is exactly the audience where competitive retention becomes interesting. These customers had both products available, paid real money for both and still disproportionately kept Claude when they cut one.

ChatGPT wins on proven retention at mass scale. Claude currently has the stronger head-to-head result among high-intent dual subscribers.

Are ChatGPT and Claude users really choosing one over the other?

A growing share of serious AI users are keeping ChatGPT and Claude together, which makes the competition much less winner-takes-all than it first appears.

The clearest business evidence comes from Ramp. Its summer analysis of more than 70,000 companies found that 52% of businesses buying either OpenAI or Anthropic were buying both.

That behavior has continued while Anthropic has pulled ahead in Ramp's AI index. In the latest update, 43.8% of U.S. businesses in the measured cohort paid Anthropic for subscriptions or tokens, compared with 39.8% for OpenAI.

Consumer data points in the same direction. YipitData finds unusually high overlap across paid AI products, especially among developers and other heavy users.

This changes how we should read churn. A company can lose part of a customer's workload without losing the customer account. Someone might keep ChatGPT while gradually moving most coding to Claude Code, or keep Claude while using ChatGPT for image generation and research.

Vendor retention can therefore look healthy even while the underlying share of work is shifting quickly.

Is Claude now harder for businesses to drop than ChatGPT?

Claude currently has stronger business adoption momentum than ChatGPT, but the evidence for genuine long-term enterprise retention is still thinner than the adoption numbers make it look.

Ramp's latest spending data puts Anthropic at 43.8% of U.S. businesses in its AI cohort, versus 39.8% for OpenAI. Earlier in the year, Anthropic was at just 16.7% in Ramp's spring report before climbing above 30% during the following quarter.

That is a huge change in a short period. It also happened while OpenAI remained widely used.

The catch is switching friction. Ramp describes AI model providers as a category with low lock-in and limited switching costs. Its finding that 52% of businesses using OpenAI or Anthropic pay both reinforces the point.

So we can say something strong about Claude's position: businesses are choosing it at an extraordinary rate, and many keep paying for it alongside alternatives.

We still cannot say with the same confidence that Anthropic has SaaS-like enterprise retention where customers are practically trapped for years. Public logo-retention and renewal-cohort data are still missing.

Claude's current stickiness seems to come mainly from people wanting to keep using the product, especially for coding and knowledge work. That is encouraging, but it is different from contractual lock-in.

Why do AI coding tools look much stickier than most AI apps?

AI coding tools are becoming some of the stickiest AI products because software development gives them a job that repeats every working day.

Developers continually open repositories, inspect files, fix bugs, write tests, review code and ship changes. An AI coding product gets another chance to prove its value every few minutes.

The best recent field evidence comes from a Microsoft study covering tens of thousands of engineers using Claude Code and GitHub Copilot CLI. The researchers found that developers who continued using command-line agents tended to be people who were actively coding, rather than a particular demographic or job level.

More importantly, agent adopters merged roughly 24% more pull requests than the researchers estimated they otherwise would have. The effect remained visible across the four-month study window instead of disappearing after the first few weeks.

Merged pull requests are an imperfect measure. They do not tell us whether the software was better, whether developers saved money or whether every company should expect a 24% improvement.

They do show something useful for retention: continued usage lined up with a task developers already perform constantly, and the measurable output effect did not quickly fade.

That makes coding one of the clearest places where AI usage is turning into habit.

Is Cursor's 250% retention number really that good?

Cursor's reported 250% net revenue retention is one of the strongest expansion numbers publicly associated with any major AI product.

The Information reported that customers who joined Cursor in March 2024 were producing roughly 250% net revenue retention a year later.

Put simply, that cohort was spending about two and a half times its original amount after both expansion and lost revenue from churn were taken into account. Traditional SaaS companies are often considered excellent when net revenue retention reaches roughly 115% to 120%.

A 250% result sits in a different league.

There is one big qualification. Net revenue retention measures money, so a smaller number of very heavy customers can drive the figure upward even when other customers leave. It tells us much more about spending expansion than about the percentage of users who stayed.

The cohort was also formed early in Cursor's rise, when its users were likely more technical and more enthusiastic than the average customer joining later.

Even after those caveats, the number is exceptional. Cursor has some of the strongest evidence we have that customers who get serious value from an AI coding tool can end up spending far more over time rather than simply renewing the same plan.

Is Claude Code becoming even stickier than Cursor?

Claude Code may be developing an even deeper daily habit than Cursor, but we still do not have a clean retention cohort that proves it.

The usage pattern is what makes Claude Code stand out. Anthropic moved Claude directly into the terminal, where developers can ask it to inspect repositories, modify files, execute commands, run tests and work through longer tasks without constantly returning to a chat window.

The Microsoft field study gives us independent evidence that this kind of command-line usage survives beyond the first experiment. Continued adoption was tied closely to actual coding activity, and the measured increase in merged pull requests persisted across the study's four-month window.

There is also an unusually strong overlap between Cursor and Claude. YipitData examined 3,891 established Cursor subscribers and found that 21.4% had paid for Claude Max, Anthropic's $100-to-$200 premium plans. Among GitHub Copilot subscribers who did not use Cursor, the equivalent figure was 11.1%. Among the comparison group of ChatGPT subscribers, it was just 4.9%.

Developers are therefore willing to pay a large premium for Claude even when they already subscribe to another coding assistant.

That tells us Claude Code is getting valuable work. It still does not tell us whether its six- or twelve-month subscriber retention beats Cursor. Anyone claiming that today is going beyond the public data.

Are developers cancelling Cursor because of Claude Code?

Claude Code is clearly taking some developer attention from Cursor, but we do not see evidence of a wholesale Cursor exodus.

The strongest clue is the amount of overlap. As seen above, more than one in five established Cursor subscribers in YipitData's cohort also paid for Claude Max, the expensive Anthropic tier.

That is a remarkable willingness to spend. Someone paying for Cursor and then adding a $100 or $200 Claude plan is probably using AI heavily enough that the tools are doing different jobs or competing for different pieces of the same workflow.

Cursor's 250% early net revenue retention also makes the “everyone is leaving Cursor for Claude” story difficult to square with the evidence. At least in that early cohort, customer spending expanded enormously.

The competitive risk is subtler. Developers can keep paying Cursor while gradually routing more high-value work through Claude Code. If that continues, Cursor could retain the subscription but lose part of the workflow.

For now, the developer market looks more like multi-tool usage than mass replacement.

Is GitHub Copilot still sticky now that Cursor and Claude Code are everywhere?

GitHub Copilot is probably still one of the stickiest AI developer products, but public data no longer gives us enough confidence to rank it above Cursor or Claude Code.

Copilot has a major advantage: distribution. GitHub already sits inside the software-development process of millions of developers and large companies. Copilot can appear inside IDEs, GitHub and enterprise workflows without asking users to adopt an entirely separate platform.

Microsoft's recent field study also included GitHub Copilot CLI alongside Claude Code and found that continued coding-agent use was tied to real coding activity. That gives us stronger evidence than simple seat counts.

But this is where the public data gets thin. Microsoft and GitHub publish plenty of adoption and activity metrics, yet clean six- or twelve-month renewal cohorts for Copilot are hard to find.

Cursor gives us a 250% net revenue retention cohort. ChatGPT gives us a large paid-consumer cohort. Claude gives us unusually revealing head-to-head subscriber behavior.

Copilot belongs in the top group because of its workflow position and distribution, but we cannot honestly call it the retention winner from the numbers available today.

Do image generators like Midjourney retain users as well as ChatGPT?

Standalone image generation appears much easier to cancel than ChatGPT for the consumers who have paid for both.

YipitData followed 426 U.S. users who were paying for Midjourney and ChatGPT during the same period and checked their subscription status roughly a year later.

Some 47.7% had dropped Midjourney while keeping ChatGPT. Only 5.2% had dropped ChatGPT while keeping Midjourney.

Among the people who ended up choosing just one, keeping ChatGPT was therefore about nine times as common.

The result makes sense when we look at how often each product can be useful. Someone may need Midjourney intensely while designing a brand, making concept art or producing a campaign, then barely touch it for several months. ChatGPT can absorb writing, research, coding, studying, planning, images and dozens of small everyday questions.

That does not mean Midjourney has poor retention among professional visual creators. This cohort is too broad to prove that.

It does show that horizontal AI has a large retention advantage when the alternative solves a narrower, more occasional job.

Are vibe-coding tools like Lovable retaining all those new customers?

Lovable has proved that vibe coding can attract paying users extremely fast, but its retention still looks much less proven than its growth.

YipitData followed 1,720 U.S. consumers who had a Lovable receipt during an earlier five-month window. In the later May-to-August measurement period, 168 showed another Lovable receipt while 1,552 did not.

We should not convert that directly into a 9.8% retention rate. Lovable's billing patterns, annual plans and receipt visibility can all distort that comparison.

The split between the two groups is more useful. Among customers with a later Lovable receipt, 56.5% were also paying for another tracked AI product. Among those without a later Lovable receipt, only 34.1% were.

In other words, Lovable currently seems stickiest among people who are already heavy AI users.

That fits the product. A casual customer may use Lovable to build one website or prototype and then leave. A founder, developer or operator building things repeatedly has many more reasons to keep it.

Vibe coding could eventually become highly retentive. The evidence today is still much weaker than what we have for ChatGPT, Claude or the best-established coding assistants.

Does paying for several AI tools mean AI retention is actually weak?

People paying for several AI tools are showing very strong attachment to AI as a category even when their loyalty to any single vendor is weaker.

Ramp found that 52% of businesses buying OpenAI or Anthropic were already paying both. YipitData sees the same behavior among consumers, particularly developers.

Poe gives us an extreme example. In YipitData's recent cohort, 46.5% of paying Poe subscribers held at least three paid AI tools, compared with just 5.2% of other paid-AI subscribers.

That creates a strange retention market. A developer can be deeply committed to AI while remaining quite willing to move individual tasks between Cursor, Claude, ChatGPT and Copilot.

The practical consequence is important. Subscription retention can overstate product loyalty. A $20 or $100 payment may survive for months after the product has stopped being the user's first choice.

The best retention evidence therefore comes from products that combine payment survival with continuing workflow usage.

What makes ChatGPT, Claude and Cursor so hard to cancel?

The stickiest AI tools today keep finding reasons to appear inside jobs people already repeat.

ChatGPT has breadth. A subscriber can arrive for writing and later use the same product for research, coding, images, files, voice or planning. That gives ChatGPT many chances to survive even when one use case becomes less important.

Claude is strongest with intensive professional work. The dual-subscriber evidence suggests that people who have tried both major assistants frequently decide Claude deserves to stay, while Claude Code gives developers reasons to use Anthropic throughout the workday.

Cursor goes even deeper into one activity. It works directly with the codebase inside an environment developers already inhabit, and its early 250% net revenue retention suggests successful customers expand usage dramatically.

GitHub Copilot benefits from an existing GitHub relationship and enterprise distribution. Products such as Notion and Canva have another advantage: their AI features live inside software, data and workflows customers already use.

The pattern is pretty clear. Retention gets stronger as the AI gains more context, sits closer to recurring work and becomes useful for more than a one-off task.

Model quality still matters a lot, but a small benchmark lead can disappear in weeks. A habit built around real work is harder to copy.

So which AI tools have the strongest retention right now?

ChatGPT has the strongest proven consumer retention at scale, Claude looks strongest among serious multi-tool AI users, and Cursor has the most extraordinary publicly reported expansion-retention number.

If we have to choose one overall winner, ChatGPT still gets it. YipitData's 40,857-person cohort showing 80.4% six-month retention gives us unusually strong evidence for a consumer AI product, and it sits alongside more than one billion weekly ChatGPT users. Very few products can combine that kind of reach with that level of paid persistence.

Claude is the most interesting challenger. Among people paying for both Claude and ChatGPT, roughly 92% still had an active Claude signal six months later. When dual subscribers ended up with just one service, Claude survived nearly five times as often as ChatGPT. Ramp's latest business data also has Anthropic ahead of OpenAI, at 43.8% versus 39.8% adoption in its U.S. cohort.

Cursor wins a different contest. Its reported 250% net revenue retention for an early customer cohort is the strongest expansion number we found among the major AI tools, although it cannot be compared directly with consumer subscriber retention.

Claude Code and GitHub Copilot also belong near the top because coding agents are showing recurring workplace usage rather than occasional visits. Microsoft's study of tens of thousands of engineers gives that claim considerably more weight than another startup ARR announcement.

Midjourney looks less durable among broad consumers when directly compared with ChatGPT. Vibe-coding products such as Lovable are growing extremely fast but still lack equally convincing retention cohorts.

So the answer is fairly sharp. ChatGPT is the safest overall retention winner. Claude may already be stickier for the most demanding AI users. Cursor has the strongest evidence that successful AI customers can dramatically deepen their spending. And AI coding tools as a group currently look closest to becoming software people feel they cannot work without.

AI tool or group	Best retention evidence we found	Current judgment
ChatGPT	80.4% six-month paid retention in a 40,857-person U.S. cohort	Strongest overall consumer retention
Claude / Claude Code	~92% of measured ChatGPT-Claude dual payers still showed paid Claude after six months	Strongest among serious multi-tool users
Cursor	~250% NRR for an early customer cohort	Strongest expansion retention
GitHub Copilot	Deep workflow distribution plus recurring usage in Microsoft's coding-agent study	Probably top-tier, but public renewal data is weaker
Midjourney	Dual subscribers were about 9x more likely to keep only ChatGPT than only Midjourney	Weaker broad-consumer retention
Lovable / vibe coding	Large acquisition, but much thinner evidence of long-term paid persistence	Too early to rank with the leaders