Can vibe coding apps make money with today's AI costs?

Are vibe coding apps already making serious money?

Vibe coding apps are now generating hundreds of millions of dollars in annual recurring revenue, with Lovable showing just how quickly people have started paying to build software with AI.

Lovable reached $400 million in annual recurring revenue in February 2026, according to figures confirmed to TechCrunch. The company had reported $100 million in July 2025, $200 million in November and $300 million in January. That amounts to a fourfold increase in roughly seven months, with another $100 million added to its annualized revenue in the final month alone.

The company achieved that scale with 146 employees, equivalent to approximately $2.74 million in annual recurring revenue per employee. For comparison, Gartner has predicted that a new generation of highly efficient software unicorns could reach $2 million per employee by 2030.

Replit and Bolt have also built paid products around natural-language software creation. Cursor serves a somewhat different audience of professional developers, but its subscription and usage-based model shows that AI-assisted programming has attracted paying customers well beyond the no-code market.

Lovable's growth establishes that customers will spend serious money on vibe coding. Whether the leading platforms keep enough of that revenue after paying their AI providers is much harder to determine. None has published the detailed inference costs and company-wide margins needed for a reliable profitability comparison.

Why does vibe coding cost so much more than normal SaaS?

Vibe coding is unusually expensive to deliver because every application change can trigger a chain of AI model calls, code inspections, tests and debugging attempts.

A conventional SaaS customer might update a database record or open a dashboard with relatively little additional computing expense. A vibe coding customer asking an AI agent to add user authentication could trigger several rounds of code generation, file reading, testing and error correction.

Replit's experience shows how large the difference can become. In July 2025, the company explained that its original Agent could work autonomously for only a couple of minutes. Its second-generation agent could operate for up to 20 minutes, with some tasks costing Replit more than $10 in computing resources.

That created a serious problem for a product initially designed around $0.25 checkpoints. Replit could no longer assume that each user request involved roughly the same amount of work.

Bolt faces another challenge as applications grow. Its documentation explains that synchronizing larger project files consumes more tokens, increasing the cost of later edits.

The most expensive customers may be those maintaining complicated applications, particularly when the agent repeatedly encounters errors. A successful five-minute request and a failed 20-minute attempt can have very different costs even when the customer asks for something that sounds equally simple.

If you want more recent data on this point, please see our latest vibe coding market report.

How much does AI actually cost to build an app?

The AI inference required to create an application can range from a few dollars to tens of dollars, depending on the model and how much work the agent performs. Published API prices let us calculate the cost of specific workloads, although they cannot establish an average cost per app.

Anthropic currently prices Claude Sonnet 5 at $2 per million input tokens and $10 per million output tokens. OpenAI charges $0.75 and $4.50 respectively for GPT-5.4 mini, while the older GPT-5 mini costs $0.25 and $2.

We compared these published prices using an illustrative application-building workload involving 10 million input tokens and 2 million output tokens across all development and debugging steps.

AI model	Input price / 1M	Output price / 1M	Cost for example workload
Claude Sonnet 4.6	$3.00	$15.00	$60.00
Claude Sonnet 5	$2.00	$10.00	$40.00
GPT-5.4	$2.50	$15.00	$55.00
GPT-5.4 mini	$0.75	$4.50	$16.50
GPT-5 mini	$0.25	$2.00	$6.50
GPT-5.4 nano	$0.20	$1.25	$4.50
GPT-5.4 mini, half the workload	$0.75	$4.50	$8.25

The workload is illustrative, not a measured average across vibe coding platforms. It excludes supplier discounts, caching, hosting, code execution and other infrastructure costs. The cheaper models also differ in capability, so they cannot necessarily complete the same task equally well.

Still, the gap is enormous. The same token volume costs $40 with Sonnet 5 and $16.50 with GPT-5.4 mini, a reduction of almost 59%.

A platform that successfully routes simple tasks to cheaper models while reserving premium models for complex debugging could save substantial money. The difficult part is making sure cheaper models do not create more failed attempts than the savings justify.

Can a $25 vibe coding subscription really cover all those AI costs?

A $25 vibe coding subscription can be profitable, but only if the platform controls how much AI work customers can consume.

Lovable's Pro plan includes 100 monthly credits for $25. Bolt charges $25 for a Pro plan starting at 10 million monthly tokens. Replit also uses credits and consumption-based charges, although its current pricing separates different types of agent usage.

These allowances cannot be compared directly. Lovable credits, Bolt tokens and Replit's agent charges represent different billing systems. None provides a simple one-to-one conversion into external API costs.

We can nevertheless test a $25 subscription using a hypothetical $1.50 per customer for payment processing and non-AI variable expenses.

With $5 of monthly inference spending, gross profit would be $18.50, representing a 74% gross margin. At $15 of inference, the margin falls to 34%. At $25, the customer generates a gross loss.

A subscriber who occasionally builds a landing page may be highly profitable, while someone running coding agents for hours can consume the entire subscription fee.

The economics become much more manageable when customers purchase extra credits after exhausting their allowance. As long as the additional charges cover the cost of serving that extra activity, frequent users can contribute revenue rather than continually eroding margins.

Can a handful of heavy users eat all the profits?

Heavy vibe coding users can consume a disproportionate share of AI spending, even when they represent a small minority of paying customers.

To understand the risk, we modeled 10,000 subscribers paying $25 per month, generating $250,000 in monthly revenue. The following distribution is hypothetical; no leading vibe coding company has publicly disclosed comparable customer-level inference data.

User group	Share of users	AI cost per user	Total AI expense
Very light	30%	$1	$3,000
Light	25%	$4	$10,000
Moderate	20%	$10	$20,000
Active	15%	$20	$30,000
Heavy	7%	$40	$28,000
Very heavy	2%	$100	$20,000
Extreme	1%	$250	$25,000

Total inference spending reaches $136,000. After another $15,000 in other variable expenses, gross profit would be $99,000, leaving a 39.6% gross margin.

The interesting result is the concentration. Just 10% of subscribers account for approximately 54% of the entire AI bill, while contributing only 10% of subscription revenue.

This is why unlimited plans are particularly dangerous when customers can launch long autonomous tasks. Pricing limits and overage charges make the expense easier to control, although platforms still need to understand how their most active customers behave.

If you want more recent data on this point, please see our latest vibe coding market report.

Are AI coding models getting cheaper fast enough?

AI coding models are getting cheaper for some comparable workloads, and today's pricing gives vibe coding platforms much more room to optimize their costs than they had with earlier premium models.

Anthropic's Sonnet 5 costs $2 per million input tokens and $10 per million output tokens, compared with $3 and $15 for Sonnet 4.6. Anthropic made the lower Sonnet 5 pricing permanent in August 2026.

OpenAI has expanded its cheaper coding-capable model lineup as well. GPT-5.4 mini costs $0.75 per million input tokens and $4.50 per million output tokens, while GPT-5.4 nano costs $0.20 and $1.25.

OpenAI reported that GPT-5.4 mini approached the larger GPT-5.4 on several coding and computer-use evaluations. That makes smaller models increasingly useful for routine parts of the development process.

However, newer models are not universally cheaper. GPT-5.4 mini costs more per token than GPT-5 mini, for example. The potential savings come from choosing the right model for each task, improving caching and reducing unnecessary work.

Another complication is Anthropic's newer tokenization system. Its documentation notes that Claude 4.7 and later models can produce approximately 30% more tokens for the same text. Comparing prices per million tokens across model generations can therefore exaggerate savings unless the workload is normalized.

The important cost measure is what the platform spends to deliver a successful application. Published token prices suggest real savings are possible, but actual task-level costs would tell us how large those savings have become.

Are more powerful AI agents making vibe coding more expensive again?

More capable coding agents can spend considerably more on a single task because they now work autonomously for much longer.

Replit provides one of the clearest documented examples. When its second-generation Agent began handling tasks lasting up to 20 minutes, certain requests cost the company more than $10.

That was a major change from the original Agent, whose short tasks had made $0.25 checkpoints commercially workable.

Replit responded in June 2025 by switching to effort-based pricing. In its July follow-up, the company reported that some simple checkpoints cost as little as $0.06, while complex ones could cost several dollars.

The pricing change also created customer frustration. Replit acknowledged that the rollout had fallen short of its standards and offered $10 in credits to affected users.

Customers want agents capable of tackling more demanding applications, but they also want predictable bills. Those two expectations are becoming harder to reconcile.

The same challenge appears in Cursor's move toward consumption-based pricing and Bolt's token allowances. As agents become capable of more autonomous work, providers have increasingly stopped treating every prompt as though it has roughly the same cost.

For now, model improvements are helping with inference efficiency, while longer and more ambitious workloads keep pushing usage upward. A platform needs both better software engineering and a billing model that can handle that variation.

Are vibe coding companies starting to earn more from usage than subscriptions?

Usage charges are becoming a meaningful additional revenue stream for vibe coding platforms, and Lovable has reported particularly strong repeat purchases.

During Stripe Sessions 2026, Lovable growth executive Elena Verna explained how the company changed its pricing while scaling from $40 million to $400 million in annual recurring revenue.

Lovable removed per-seat pricing, introduced credit top-ups and allowed customers to pay for additional usage rather than forcing them into progressively larger subscriptions.

The company discovered that many of its best customers preferred purchasing extra credits. According to Verna, top-up repurchase rates were at least as high as subscription renewal rates.

Customers who repeatedly buy more credits have shown a willingness to pay for additional development work, rather than treating the monthly subscription as an unlimited entitlement.

Lovable also reported that allowing top-ups caused some subscription revenue to shift into consumption-based purchases, while overall collections continued growing. The company did not disclose enough data to calculate the precise percentage of revenue from top-ups.

We can see a similar direction in Replit's charges for agent effort. Recurring subscriptions and additional consumption can work alongside each other, rather than forcing a company to choose between the two.

The remaining question is how much gross profit those extra purchases generate. Repeat spending is encouraging, but profitable incremental usage requires the platform's charges to exceed the actual cost of serving it.

If you want more recent data on this point, please see our latest vibe coding market report.

Is Lovable's free plan worth the AI computing bill?

Lovable's free plan appears to be an effective way to turn new users into paying customers, although the company has not disclosed the inference spending needed to calculate its acquisition economics.

Lovable has reported double-digit conversion rates from free users. Stripe compared that performance with ChartMogul research showing an 8% median conversion rate for B2B software products.

The comparison is encouraging, although conversion definitions and measurement periods may differ.

A separate Lovable partnership with Lenny's Newsletter converted more than 40% of participating users, according to Stripe. That was a specific promotional cohort and should not be mistaken for Lovable's typical conversion rate.

There is also evidence that free access creates substantial activity. TechCrunch reported that Lovable's SheBuilds promotion, which made the platform free for one day in March 2026, generated more than 500,000 projects built or updated. Normal daily activity was approximately 200,000 projects.

That represents about 2.5 times the usual activity during the promotion. It demonstrates how responsive usage can be to eliminating the initial price barrier, although it tells us nothing about how many of those projects later generated revenue.

Free plans can work particularly well for software creation because customers can experience the product before deciding whether the result is worth paying for.

Still, without cohort-level acquisition spending, AI costs and conversion revenue, we cannot say that Lovable's free users are profitable on average. The reported conversion performance makes the strategy credible, but it does not settle the economics.

Does Lovable's huge revenue mean the company is profitable?

Lovable's reported revenue gives it considerable financial scale, but the company's actual profitability remains unverified.

The company has not publicly released a detailed income statement showing model inference costs, cloud expenses, gross profit and operating income.

Its small employee base is impressive, but AI infrastructure can remain expensive even when payroll is relatively low. Recurring revenue also measures a different thing from profit.

To see the financial sensitivity, consider a business operating at $400 million in annualized revenue.

At a 50% gross margin, annualized gross profit would be $200 million. At 70%, it would reach $280 million. At 80%, it would be $320 million.

That's a $120 million difference between the first and last scenario, before salaries, marketing, research and other operating expenses.

These are mathematical scenarios rather than estimates of Lovable's financial results. The company could also have very different margins across subscriptions, credit top-ups and cloud services.

One additional distinction is important: $400 million in ARR describes an annualized recurring-revenue run rate, rather than the revenue necessarily recognized during the preceding year.

Lovable has clearly established the scale to support a substantial business. But calling it profitable would require knowing how much it spends to generate and retain that revenue.

Can vibe coding apps keep making money after customers finish building?

Vibe coding platforms face a retention problem because some customers only need an AI builder for a few weeks, but hosting and ongoing maintenance can give them reasons to keep paying.

Someone creating a simple website might finish the project, export the code and cancel. That is a very different relationship from using accounting software or a CRM every working day.

Lovable and Replit have responded by offering more than application generation.

Lovable Cloud provides backend infrastructure for applications built on the platform. Lovable AI allows customers to add AI features to those applications, with usage-based billing for the services. Replit combines its development environment with application deployment and hosting.

Bolt also includes web hosting, custom domains and databases in its paid plans.

Hosting creates another source of revenue after the initial build. A customer whose application receives regular traffic may continue paying for compute, storage, database usage and other services.

There is a limit to this opportunity, however. Vercel, Netlify and major cloud platforms already offer competing infrastructure. Customers who can export their projects may decide to host them elsewhere.

Lovable's published ecosystem usage gives some indication of the opportunity. Stripe reported more than 300 million monthly visits to applications and websites built with Lovable. That traffic is not necessarily hosted or monetized by Lovable, so we cannot convert it directly into infrastructure revenue.

The strongest long-term customer is likely to be someone who continues improving and operating an application. For those customers, vibe coding can become an ongoing software expense rather than a one-time development purchase.

If you want more recent data on this point, please see our latest vibe coding market report.

Can business customers make vibe coding much more profitable?

Business customers give vibe coding platforms a way to charge more for security, administration and collaboration without necessarily increasing AI usage by the same proportion.

Lovable's Business plan starts at $50 per month, compared with $25 for Pro. Both start with 100 monthly credits, but Business includes additional administrative and security functionality.

The distinction is commercially attractive. An organization might pay twice as much for better access controls and workspace management while consuming the same included development credits.

Lovable has also attracted established businesses. TechCrunch identified Klarna and HubSpot among its enterprise users.

In June 2026, Forbes reported that enterprise customers represented approximately $20 million in ARR, citing a source familiar with Lovable's business. That suggests enterprise contracts were still a relatively small part of its total recurring revenue at the time.

Business customers can nevertheless be more expensive to support. Large organizations may require stronger security, service guarantees, custom integrations and more complicated development workflows.

We cannot conclude that enterprise contracts have higher gross margins without their actual costs.

Even so, enterprise functionality gives vibe coding platforms more ways to charge for something customers value beyond model tokens. This is particularly useful as basic AI code generation becomes more widely available.

Could OpenAI, Anthropic and Google take over the vibe coding market?

OpenAI, Anthropic and Google are credible threats to independent vibe coding startups because they control important AI models and increasingly offer coding tools of their own.

OpenAI has Codex, Anthropic has Claude Code, and Google's Gemini ecosystem includes AI-assisted development products. These companies can place increasingly capable coding agents directly in front of existing customers.

That puts startups such as Lovable and Replit in an awkward position. They often depend on the same model providers that could compete for their users.

There is still a meaningful difference between generating code and delivering an application that a non-developer can use, publish and maintain.

Lovable emphasizes visual development and accessible app creation. Replit combines coding, execution and deployment in one environment. Bolt lets customers build and publish applications through an integrated interface.

Those workflows involve more than access to a language model.

At the same time, model providers earn API revenue when independent companies use their systems. Supporting third-party application platforms can therefore remain commercially attractive even while they develop competing products.

The risk is greatest for startups whose main advantage is providing a simpler interface to a widely available coding model. That experience can be copied relatively easily.

Companies that develop strong distribution, reliable application management and useful integrations have more ways to remain competitive. Whether those advantages will be enough against major AI labs is still unsettled.

What happens to vibe coding profits if AI costs fall another 50%?

A further 50% reduction in AI inference costs could transform the gross margins of expensive vibe coding subscriptions, especially for companies that already control usage.

We modeled a $25 monthly subscriber with different inference expenses, assuming $2 in other variable costs. For each scenario, we cut inference costs in half while leaving the customer's workload and subscription price unchanged.

Current AI cost	Other costs	Current gross margin	Margin after 50% AI cut
$2	$2	84%	88%
$5	$2	72%	82%
$8	$2	60%	76%
$10	$2	52%	72%
$15	$2	32%	62%
$20	$2	12%	52%
$30	$2	-28%	32%

The biggest relative gains appear among customers whose current AI consumption is expensive.

A subscription producing a 12% gross margin with $20 of monthly inference costs would reach 52% if that expense fell to $10. Even the loss-making $30 inference scenario becomes gross-margin positive.

These calculations assume the company keeps all the savings. In reality, lower costs may encourage competitors to offer larger allowances, reduce prices or introduce more advanced features.

Users may also consume more AI as agents become capable of longer tasks.

The numbers show that a vibe coding platform doesn't necessarily need dramatic subscription price increases to achieve healthy gross margins. It can get there through cheaper inference, better task routing and consumption limits.

Which vibe coding business model has the best chance of making money?

The best-supported vibe coding business model today combines a subscription, limits on included AI usage, paid overages and recurring services such as hosting.

Lovable, Replit and Bolt have approached the problem differently, but their pricing changes point in the same direction.

Lovable uses workspace subscriptions and credits, with additional purchases and cloud billing. Replit charges for the amount of effort performed by its agent. Bolt offers defined monthly token allowances alongside hosting and database capabilities.

These models give platforms ways to recover their expenses from customers who use more resources.

A flat unlimited subscription is much harder to manage. As Replit's experience demonstrated, one advanced coding session can cost far more than the small fixed charge originally associated with an individual request.

The economics also improve when companies sell services that customers continue using after development. Hosting, enterprise security and application management can generate revenue independently of new coding sessions.

There is a trade-off: more sophisticated billing creates complexity for customers. Replit's initial effort-based rollout showed that unexpected charges can lead to frustration, even when the pricing better reflects computing costs.

For now, the strongest commercial approach appears to be a reasonably priced entry subscription, clear spending limits and additional charges for customers who need substantial computing power.

This gives casual users an accessible way to start while allowing the platform to earn more from demanding projects.

If you want more recent data on this point, please see our latest vibe coding market report.

Can vibe coding apps make money with today's AI costs?

Yes. Vibe coding apps can make money with today's AI costs, and the combination of cheaper inference, controlled usage and recurring platform services makes the business increasingly credible. But the leading startups have yet to publicly demonstrate consistent company-wide profitability.

Commercial demand is real. The rapid growth of Lovable and the development of paid products from Replit, Bolt and other companies show that customers are willing to spend money building software through AI.

The original pricing problem is also being addressed. As seen above, multiple providers have independently moved toward charging for consumption, restricting included usage or billing according to the effort required.

That's particularly important because the cost of AI coding varies enormously between customers and tasks. An occasional user can be inexpensive to serve, while an autonomous agent working through complicated debugging sessions may generate substantial inference costs.

There is also a clear financial advantage from using cheaper models where their capabilities are sufficient. Our pricing calculations show that the same token workload can cost several times more with a premium model than with a smaller alternative. Platforms that reduce unnecessary context and failed attempts have another opportunity to save.

The remaining uncertainty concerns actual margins. Lovable's revenue and small employee base are encouraging, but without disclosed inference expenses, hosting costs and customer acquisition spending, profitability claims would go beyond the available evidence.

Looking further ahead, the biggest opportunity may come from keeping customers after they finish building. Hosting, maintenance, collaboration and enterprise features can make the relationship more durable while reducing dependence on revenue from individual coding sessions.

Our conclusion is that vibe coding already has the ingredients of a viable software business. The companies most likely to make attractive profits will be those that charge appropriately for expensive AI work and keep earning revenue from the applications their customers build.

Cheap AI will help, but the winners will still need to turn that computing advantage into reliable software, satisfied customers and recurring revenue.