Which AI wrappers are still growing now?

Why did everyone think AI wrappers would die?

The fear that AI wrappers would disappear made sense for simple GPT front ends, but it badly underestimated how valuable the application layer could become.

When ChatGPT exploded, thousands of startups could suddenly sell summarization, copywriting, research, coding or image generation by putting an interface around somebody else’s model. The weakness was obvious. OpenAI, Anthropic or Google could add the same feature to their own products, while another startup could copy the interface in a weekend.

That threat has become more serious. ChatGPT has expanded into browsing, deep research, coding, image creation and agentic work. Claude has moved aggressively into coding and computer use. Gemini reaches users through Google’s existing products. Entire startup categories have watched features that once looked differentiated turn into standard model capabilities.

Yet the application layer did not disappear. Cursor, Lovable, Replit, Harvey, Legora, Sierra, Perplexity and Genspark have all reached very large revenue levels while relying heavily, or historically relying heavily, on outside foundation models.

What changed is what customers pay them for. Cursor sits inside software development. Harvey and Legora sit inside legal work. Sierra handles customer operations. Lovable and Replit build working applications. Gamma creates finished presentations and websites.

Access to an intelligent model has become cheap and widely available. Turning that intelligence into a reliable piece of work still has a lot of value.

Which AI wrappers are actually still growing fast now?

Cursor, Lovable, Replit, Harvey, Legora, Sierra, Perplexity and Genspark have the clearest evidence of major commercial growth today.

Cursor is in a league of its own on absolute scale. Forbes reported that its annualized revenue reached roughly $4 billion in June, after crossing $3 billion in late April and $2 billion in February.

Lovable reported $500 million in annualized revenue by June, up from about $200 million in November. Replit was estimated by Sacra at roughly $525 million in April, versus around $300 million at the end of last year.

Perplexity crossed $450 million in recurring revenue according to the Financial Times. Harvey has now passed $400 million ARR. Anthropic says Genspark has exceeded $250 million ARR since moving to its Super Agent product. Sierra reported more than $150 million ARR, while Legora crossed $100 million only 18 months after launch.

Those figures are not perfectly comparable. Some are true ARR, some are annualized recent sales, and Replit’s latest figure comes from an outside estimate. Still, the order of magnitude is impossible to ignore. We are looking at several application companies adding hundreds of millions of dollars in revenue while the underlying models become easier to access.

Company	Latest useful revenue signal	Main AI workflow
Cursor	~$4B annualized revenue	Software development
Replit	~$525M estimated annualized revenue	Software creation
Lovable	$500M+ annualized revenue	App building
Perplexity	$450M+ recurring revenue	Search and agents
Harvey	$400M+ ARR	Legal work
Genspark	$250M+ ARR	General knowledge work
Sierra	$150M+ ARR	Customer service
Legora	$100M+ ARR	Legal work

Can we trust these huge AI wrapper revenue numbers?

We can trust the direction much more than the exact ranking, because AI companies increasingly use ARR, annualized revenue and run rate to describe different things.

Traditional SaaS ARR usually refers to contracted recurring subscription revenue. In AI, a company may take one strong month of consumption and multiply it by twelve. Usage-based revenue can swing quickly, consumer subscriptions can churn, and rapid adoption can make a young company’s run rate look more settled than it really is.

The problem became especially visible when Cluely’s CEO publicly retracted revenue figures he had previously given a reporter. Investors have also become more vocal about startups stretching the definition of ARR.

That is why we give much more weight to a sequence than to one impressive announcement.

Cursor moved through roughly $1 billion, $2 billion, $3 billion and $4 billion in successive public reports. Lovable went through approximately $100 million, $200 million, $400 million and $500 million. Harvey passed $100 million and has since moved beyond $400 million. Those repeated increases are harder to explain away as one unusually strong month.

We also look for another metric beside revenue. Lovable now hosts around 60 million projects that attract roughly 900 million visits per month. Harvey serves more than 3,000 customers and 80% of the Am Law 100. Perplexity has more than 100 million monthly users. Legora serves more than 1,000 firms.

So the exact leaderboard deserves caution. The conclusion that several AI application companies are still growing very quickly does not.

Revenue claim	How useful it is
Repeated ARR milestones plus customer growth	Very strong
Contracted enterprise ARR	Strong
Company-reported subscription ARR	Strong
Annualized recent revenue	Useful, with caution
Outside revenue estimate	Directional
One founder run-rate claim	Weak without confirmation

Is Cursor still growing even with Claude Code everywhere?

Yes. Cursor is still growing extremely fast, and the latest numbers make the idea that Claude Code has already killed it look premature.

The interesting part is what happened after the backlash began. Developers were publicly talking about switching from Cursor to Claude Code, OpenAI was pushing Codex harder, and some investors argued that the coding layer would eventually collapse into the model providers.

Meanwhile, Cursor’s annualized revenue kept rising.

Bloomberg reported roughly $2 billion in February, double the level three months earlier. Reporting in April put the figure around $3 billion. Forbes then reported approximately $4 billion in June.

That means Cursor added roughly $2 billion of annualized revenue in about four months after already becoming a multibillion-dollar business.

The customer mix helps explain it. Bloomberg reported earlier this year that large companies accounted for around 60% of Cursor’s revenue. Some individual developers may prefer Claude Code, but higher-spending enterprise teams have been moving in the opposite direction.

Cursor is also building more of the technology itself. Its Composer model helped improve product economics, and the company has been investing heavily in coding-model research.

So Cursor is becoming harder to describe as a conventional wrapper. It still benefits enormously from external models, but the editor, repository context, agent system, enterprise relationships and increasingly its own models all sit between the foundation model and the customer.

Can Lovable and Replit keep growing after the vibe-coding boom?

Yes. Lovable and Replit are still growing quickly enough that vibe coding has already moved beyond a one-season consumer craze.

Lovable reached about $100 million in annualized revenue last summer, around $200 million by November, $400 million by February and more than $500 million by June. Its latest funding round valued the company at $13.3 billion, twice the valuation from its previous round.

The usage underneath that revenue is still expanding. Lovable says users create around one million new projects every week. By its latest fundraising announcement, the platform hosted roughly 60 million projects attracting about 900 million monthly visits.

Growth has slowed from the insane early phase, which was inevitable. Going from $200 million to $400 million happened faster than moving from $400 million to $500 million. Lovable was also expecting to finish August around $600 million annualized revenue, according to reporting around its latest financing, so the business was still adding meaningful revenue after crossing half a billion.

Replit has taken a different route. The company disclosed that annualized revenue had jumped from $2.8 million to $150 million in less than a year. Sacra later estimated around $300 million at the end of last year and $525 million by April.

Replit is also widening the product beyond programmers. Agent can build complete applications, while newer products move into presentations, videos, design and other white-collar work. The company has more than 60 million users and has been pushing further into enterprise distribution through large cloud marketplaces.

Lovable currently has the cleaner public revenue trail. Replit has the broader installed base and a longer history as a development platform.

Either way, people are paying hundreds of millions of dollars to generate working software through an application layer built on top of foundation models. That market is very real now.

Are Harvey and Legora proving that vertical AI wrappers work?

Yes. Harvey and Legora make legal AI the strongest current example of a vertical AI application becoming a serious software category.

Harvey passed $100 million ARR last year. The company has now crossed $400 million, according to its latest financing announcement, after adding more than $100 million of ARR in a single quarter earlier this year.

Its customer footprint has expanded with the revenue. Harvey now serves more than 3,000 customers, including 80% of the Am Law 100, about 20% of the Fortune 500 and legal teams at half of the Fortune 10.

Legora gives us the second data point needed to show that this is bigger than one exceptional company. Legora crossed $100 million ARR only 18 months after launching publicly. It now serves more than 1,000 law firms and other legal organizations across dozens of markets.

Both companies are moving deeper into legal work rather than simply offering lawyers a specialized chat window. They connect to firm documents, institutional knowledge, permissions, workflows and external legal sources. Harvey has also released its own post-trained open-weight model and is putting more money into model training.

Law firms could already buy ChatGPT or Claude directly. Many are paying Harvey or Legora anyway.

That tells us where the value sits: the winning legal products understand how lawyers actually work and can fit AI into that environment without asking every lawyer to assemble the workflow themselves.

Metric	Harvey	Legora
Current ARR milestone	$400M+	$100M+
Customers	3,000+	1,000+ firms
Adoption signal	80% of Am Law 100	Major global law firms across dozens of markets
Product direction	Legal agents, workflows, own post-trained models	Legal workflows, integrations, firm knowledge

Is Sierra still growing in customer-service AI?

Yes. Sierra has already passed $150 million ARR and continues to push deeper into long-running customer-service work rather than simple support chat.

Sierra reached $100 million ARR in only seven quarters and reported more than $150 million shortly afterward. Its customer base includes large companies such as ADT, Cigna, Rivian, SoFi and SiriusXM, with a substantial share of customers generating billions of dollars of annual revenue.

The more recent product releases show where Sierra wants the business to go next. Its agents can maintain context across customer interactions lasting days or weeks, and Sierra charges around outcomes rather than simply charging for tokens or model access.

That pricing choice says a lot about the product. A retailer does not particularly care which language model answered a question. It cares whether the return was completed, the account was fixed or the customer problem was resolved.

Sierra becomes harder to replace as more of that operational work passes through the platform.

That is a much stronger position than selling companies a prettier chatbot.

Are Perplexity and Genspark still growing as ChatGPT gets more agentic?

Yes. Perplexity and Genspark are still growing, although they face more direct pressure from ChatGPT, Claude and Gemini than the vertical AI companies do.

Perplexity is the bigger business today. The Financial Times reported that recurring revenue passed $450 million after monthly revenue jumped around 50% during the company’s move from AI search toward agents.

Perplexity also has more than 100 million monthly users and tens of thousands of enterprise customers. Its newer products cover browser activity, research and longer multi-step tasks, with subscriptions ranging from mainstream consumer plans to a $200 premium tier.

Genspark has followed an even more aggressive agent strategy. Anthropic’s recent case study says Genspark has passed $250 million ARR since pivoting to its Super Agent product. The system can coordinate more than 150 specialized tools to produce slides, spreadsheets, documents, websites and other work.

The risk is obvious for both companies. ChatGPT, Claude and Gemini increasingly perform the same broad research and execution tasks.

Their answer has been to move faster into orchestration. Perplexity routes tasks across several models and products. Genspark coordinates a large toolset around Claude and other infrastructure.

This is a tougher battlefield than legal AI because the foundation-model companies want the same general-purpose user relationship. For now, though, both businesses are still adding enough usage and revenue to remain credible independent products.

Is Gamma still growing, or has AI presentation growth flattened?

Gamma is still gaining users quickly, but its current revenue growth is harder to measure because the company has not published a newer ARR milestone than $100 million.

Gamma reported roughly $100 million ARR and 70 million users while already profitable late last year. TechCrunch said the service was approaching 100 million users a few months later, and Gamma has since passed that 100 million-user mark.

That is roughly 30 million additional registered users after reaching $100 million ARR.

The missing piece is a refreshed revenue number. Continued user growth does not automatically mean ARR is rising at the same speed, and repeating the old $100 million figure would give us false precision about what is happening today.

Still, Gamma has cleared several tests that most consumer AI wrappers never do. The product reached huge distribution, found paid demand and did it while remaining profitable. It also expanded beyond AI presentations into websites, social posts, images and broader visual communication.

Gamma therefore belongs among the durable AI application companies. We just have less evidence that its revenue is accelerating right now than we have for Cursor, Lovable or Harvey.

Is Fyxer still growing, or did its AI email boom already cool down?

Fyxer is still growing, but its latest numbers show a major slowdown from the company’s initial breakout.

The AI executive assistant went from roughly $1 million in annualized revenue at the end of 2024 to about $30 million by the end of 2025. Sacra estimates it reached around $35 million in February.

Those two periods tell very different stories. The first represents roughly 30-fold growth. The next publicly visible step added around $5 million.

Fyxer still has a useful product position. It works inside Gmail, Outlook and calendars, drafts replies, organizes inboxes and produces meeting notes. Users do not need to remember to visit a separate AI chatbot every time they want value from it.

The company also said its system processed around 1.4 billion emails during 2025, so there is real activity underneath the subscriptions.

Fyxer remains a good example of a smaller workflow wrapper that found product-market fit. Based on the latest revenue evidence, we would no longer put it in the same growth tier as Cursor, Lovable, Harvey or Genspark.

Are thin AI wrappers still growing?

Thin AI wrappers can still make money, but the biggest current growth is concentrated in products that do much more than send prompts to a model.

Look at what the leaders actually deliver.

Cursor edits a real codebase. Lovable and Replit build and deploy software. Harvey and Legora work across legal documents and firm knowledge. Sierra completes customer-service actions. Gamma produces finished visual material. Genspark can coordinate more than 150 tools.

The products are getting thicker because simple interfaces are getting easier to reproduce.

A standalone AI summarizer had a real advantage when general chatbots were bad at files and long context. That advantage gets smaller whenever the base models improve. The same model improvement can help a deeply integrated application because the better intelligence flows into an existing workflow.

That creates an important split in the market today.

Generic wrappers have less room to hide. Workflow products can actually become better when OpenAI, Anthropic or Google ships a stronger model.

Why haven't ChatGPT and Claude killed these AI wrappers?

ChatGPT and Claude have absorbed plenty of AI features, but recreating an entire customer workflow is much harder than recreating one feature.

Coding shows the difference clearly. Anthropic can release a brilliant coding model and Claude Code can win developers. Cursor still has an editor, enterprise deployments, repository context, its own model work and teams that have already built their development habits around it.

The same thing happens in legal AI. Claude can become much better at reading contracts without suddenly inheriting Harvey’s connections to law-firm systems, customer relationships, permissions and institutional knowledge.

Foundation-model companies also compete fiercely with each other. Application companies can benefit from that.

Perplexity can route work between models. Gamma uses several models behind one interface. Coding products can change model suppliers or offer several choices. Cheaper and better inference therefore lowers an application company’s input cost while improving what the product can do.

There is still a real platform risk. OpenAI, Anthropic and Google will keep moving into applications, and some startups will get squeezed.

But a new ChatGPT feature does not automatically move an enterprise workflow, its data and its users to OpenAI.

Do growing AI wrappers eventually need their own models?

The strongest AI wrappers are increasingly building specialized models, but they do not need to recreate GPT or Claude from scratch.

Cursor is investing heavily in coding models and has already shipped its own Composer model. Harvey recently released Tenet, a post-trained open-weight legal model. Lovable now offers an internally trained model alongside frontier-model options.

There is a straightforward economic reason for this.

Once a company serves enough traffic, using a frontier model for every tiny task can become unnecessarily expensive. A specialized internal model may be faster or cheaper for predictable work, while Claude, GPT or Gemini handles the difficult tasks.

Specialization can also improve quality. Harvey has access to legal workflows and feedback that a general model provider cannot optimize around as narrowly. Cursor sees enormous volumes of real coding interactions.

The likely endpoint is a mixed stack. Application companies can rent general intelligence, train narrow intelligence where it gives them an advantage and control the orchestration layer in between.

That is already happening across several of the largest wrapper-origin businesses.

What makes an AI wrapper hard to replace now?

The AI wrappers that are hardest to replace own the workflow, the context or the finished output that users care about.

Repeated use helps first. A developer can spend hours every day inside Cursor. A lawyer can run active matters through Harvey. Sierra can handle customer conversations continuously. Products used several times a day get much more opportunity to become habitual than one-purpose AI generators.

Context comes next. Codebases, company documents, customer histories, preferences, integrations and previous work all make switching less frictionless.

Then there is execution. Generating an answer is becoming cheap. Actually changing a repository, deploying an application, completing a return, preparing legal work or publishing a presentation requires a larger system around the model.

Distribution can be just as powerful. Lovable says the applications hosted on its platform receive around 900 million visits every month. Gamma spread through shareable presentations and documents. Harvey and Legora expand firm by firm. Replit already had a huge developer audience before the current AI boom.

The moat therefore looks much more like conventional software than early AI hype suggested: workflow depth, customer data, habit, distribution and integrations, with AI making the product substantially better.

Which AI wrapper categories are strongest right now?

AI coding and legal AI currently have the strongest evidence because each category has produced several independent companies with large and still-rising revenue.

Coding gives us Cursor at roughly $4 billion annualized revenue, Lovable above $500 million and Replit estimated above $500 million. These products target different users, which makes the pattern more convincing. Cursor focuses heavily on professional developers, while Lovable and Replit have pushed software creation toward people who may barely code at all.

Legal AI has two clear leaders. Harvey has crossed $400 million ARR while Legora passed $100 million in only 18 months. Both sell into law firms that already have easy access to general-purpose AI products, yet those firms are still buying specialized software.

Customer-service agents look strong too. Sierra passed $150 million ARR and is embedded in large enterprises.

General agents remain commercially impressive but more exposed. Perplexity and Genspark have hundreds of millions in recurring or annualized revenue, while OpenAI, Anthropic and Google are all chasing the same broad knowledge-work territory.

Creative productivity has already produced a profitable $100 million business in Gamma, although we currently have less visibility into how quickly revenue is growing beyond that level.

Our current ranking is therefore fairly clear: coding first, legal AI second, customer-service agents next, with broad agents carrying more platform risk despite their rapid growth.

Which AI wrappers are still growing now?

The clearest answer today is Cursor, Lovable, Replit, Harvey, Perplexity, Genspark, Sierra and Legora, while Gamma is still expanding strongly in users and Fyxer is growing at a much slower pace than during its breakout.

Cursor is the standout. Its reported annualized revenue has gone from roughly $2 billion to $4 billion in only a few months. Lovable has crossed $500 million and continues adding projects at enormous scale. Replit is estimated around the same revenue level. Harvey has moved beyond $400 million ARR. Perplexity exceeds $450 million in recurring revenue. Genspark has passed $250 million ARR. Sierra and Legora have cleared $150 million and $100 million respectively.

The same pattern keeps appearing across very different categories. The products growing fastest today sit inside a job people perform repeatedly: coding, legal work, customer service, research or content creation.

That is also why the word “wrapper” has become less useful.

The weak version of the business still exists: take a foundation model, add a narrow interface and hope users keep paying. Model providers can erase that advantage very quickly.

The companies still compounding have gone much further. They control context, tools, integrations and increasingly parts of the model stack itself. Most importantly, they help users finish something rather than merely generate an answer.

So yes, AI wrappers are still growing—some at extraordinary speed. But the winners are gradually turning themselves into full software platforms before the underlying models can turn them into features.