Are any GPT wrappers still making a lot of money?

What actually counts as a GPT wrapper today?

A GPT wrapper today is best understood as an AI application that still depends heavily on somebody else’s foundation models, while adding its own interface, workflow, data or tools around them.

The definition has become messy because the successful companies have changed. In the early ChatGPT period, a wrapper often meant little more than an OpenAI API call behind a nicer interface. A copywriting tool might send a prompt to GPT-3, return the result and charge $20 a month.

Cursor, Harvey, Lovable and Gamma all depend on outside AI models to varying degrees, so people still casually call them wrappers. Yet the products now contain far more than the model call. Cursor controls much of the coding environment around the model. Harvey connects AI to legal documents, permissions and firm workflows. Lovable takes a natural-language request all the way toward a working application. Gamma turns generated content into editable presentations, websites and documents.

The Wall Street Journal has recently used the term “model harness” for this layer: the software that supplies models with context, memory, tools and actions. That describes the stronger application companies better than the old idea of a thin wrapper.

For this analysis, we will keep the definition fairly broad. If an application still depends materially on OpenAI, Anthropic, Google or another outside model provider, it belongs in the discussion. We will then separate the thin wrappers from the companies that have built much more around those models.

Why did people start saying GPT wrappers were dead?

The “GPT wrappers are dead” argument came from a real wave of failures and slowdowns after ChatGPT absorbed jobs that startups had previously sold as standalone products.

Jasper is the obvious example. The company became one of the first generative-AI software stars by packaging GPT-powered writing into a much easier product for marketers. The Information reported that Jasper reached around $80 million in annual recurring revenue by the end of 2022 and roughly $90 million soon after.

Then ChatGPT arrived.

Suddenly, a user who wanted a blog draft, product description or marketing rewrite could ask OpenAI directly. Jasper subsequently cut its revenue projections by at least 30%, reduced staff and shifted its focus toward larger marketing teams. Its internal share valuation was also marked down by about 20%.

That experience gave the wrapper argument real credibility. When most of the product value comes from making a foundation model easier to access, the model company can destroy that advantage very quickly by improving its own interface.

A lot of early AI writing, summarization and document-chat products faced the same pressure. The mistake was extending that lesson to every software company using outside AI models.

Are any GPT wrappers still doing more than $100 million a year?

Yes, several GPT-wrapper-style companies are currently well above $100 million in annualized revenue, and the biggest one is measured in billions rather than millions.

Forbes recently reported that Cursor passed $4 billion in annualized revenue. Lovable crossed $500 million. The Information puts Harvey above $350 million. Recent Financial Times reporting puts Legora around $150 million. Gamma has crossed $100 million in ARR.

Those five companies alone add up to more than $5.1 billion in reported annualized revenue.

We should be careful with the accounting language. These are private companies, and “annualized revenue” can simply multiply a recent month or quarter into a yearly figure. The Financial Times recently highlighted this problem across fast-growing AI companies: an annualized run rate can make volatile usage revenue look more recurring than it really is.

Even after allowing for that, the scale is too large to dismiss. Cursor could disappear from the calculation entirely and the remaining four companies would still represent more than $1 billion of annualized revenue.

The interesting part is that none of these winners looks much like the classic 2023 wrapper anymore.

Company	Recent reported revenue scale	What customers are actually buying
Cursor	$4B+ annualized	AI coding environment and agents
Lovable	$500M+ annualized	End-to-end app creation
Harvey	$350M+ annualized	Legal AI workflows
Legora	~$150M ARR	Legal workspace and agents
Gamma	$100M+ ARR	AI presentations, documents and websites

Are the biggest GPT wrappers still growing right now?

Yes, the biggest model-dependent AI applications are still growing extremely fast today, which makes it hard to describe their revenue as leftover momentum from the first ChatGPT boom.

Cursor provides the clearest trajectory. Forbes reported roughly $2 billion in annualized revenue in February, $3 billion by late April and more than $4 billion a few weeks later. Around three quarters of the run rate was coming from business customers, so the growth had already moved far beyond individual developers buying cheap subscriptions.

Lovable has followed a similar curve at a smaller scale. The company reported roughly $400 million in annual recurring revenue in February and more than $500 million by June. Its latest funding announcement also disclosed that roughly 60 million projects are now hosted on the platform and attract around 900 million visits per month.

Legal AI is moving quickly too. Harvey is currently above $350 million in annualized revenue, according to The Information, up more than 80% from roughly $190 million at the beginning of the year. Legora has reached around $150 million ARR after being closer to $50 million at the end of last year. The Financial Times recently reported more than seven consecutive quarters of at least 50% ARR growth for Legora.

Three separate categories, coding, app creation and legal AI, are producing unusually fast application-layer growth at the same time. That is much harder to explain as one anomalous company.

Is Cursor still basically a GPT wrapper?

Cursor still depends heavily on outside frontier models, but calling the current company a thin GPT wrapper would miss most of what developers are paying for.

Cursor started with an easy-to-understand idea: put powerful AI models inside a code editor. A developer could get much of the underlying intelligence from OpenAI or Anthropic directly.

The product has since moved much deeper into software development. Cursor indexes repositories, decides which code should enter the model context, edits multiple files, runs agents, executes longer coding tasks and lets developers switch between different model providers. Cursor also develops some models itself.

That surrounding product has translated into extraordinary commercial scale. Forbes recently put Cursor above $4 billion in annualized revenue, twice its reported run rate only a few months earlier. SpaceX subsequently agreed to acquire its parent company Anysphere in a transaction valuing the business at roughly $60 billion.

The important part for the wrapper debate is how Cursor reacted when Anthropic launched Claude Code and OpenAI pushed deeper into coding. Cursor kept growing while those suppliers became direct competitors.

Developers apparently value more than access to the smartest model. They also value where the model operates, what context it sees, how easily it can change a repository and how naturally AI fits into the rest of the coding session.

Is Lovable still basically a GPT wrapper?

Lovable currently looks much more like an application-building platform than a simple interface around a coding model.

The original pitch sounded extremely wrapper-like: describe the application you want and AI generates the code. Anyone watching OpenAI, Anthropic or Google improve their coding models could reasonably wonder what Lovable would still contribute.

Users, however, increasingly ask Lovable for the finished result rather than raw generated code. The platform handles much of the project structure, iteration, integrations, debugging and deployment around that generation step. That opens app development to customers who might have no idea what to do with several thousand lines of AI-generated React code.

The numbers are substantial. TechCrunch reported more than $500 million in annualized revenue and roughly one million new projects being created each week earlier this summer. More recently, Lovable said it hosts around 60 million projects, and investors valued the company at $13.3 billion in a new $400 million round.

Its infrastructure choices are also changing. Lovable signed a multiyear Google Cloud agreement that was expected to increase its usage fivefold, while the company has also started offering an internally trained model alongside outside frontier models.

That combination tells us more than the valuation does. Lovable is happy to buy intelligence from Google and other suppliers while gradually owning more of the surrounding system itself.

Are legal AI wrappers like Harvey and Legora really making serious money?

Yes, Harvey and Legora are currently making hundreds of millions of dollars from legal AI while still relying heavily on foundation models built elsewhere.

Legal AI gives us an unusually clean test because the Financial Times recently described companies such as Legora as customized wrappers around models from OpenAI, Anthropic and other labs.

Harvey is now generating more than $350 million in annualized revenue, according to The Information. The company deliberately uses several model providers, including OpenAI, Anthropic and Google, rather than tying the whole product to one laboratory.

Legora has reached roughly $150 million ARR. Its customer base has also climbed to around 1,500 law firms and in-house legal teams, while the company says more than 100,000 legal professionals use the platform. Recent Financial Times reporting says Legora has posted at least 50% ARR growth for seven consecutive quarters.

Law firms could already use ChatGPT or Claude for drafting and summarization. They keep paying Harvey and Legora because professional legal work involves much more: matter permissions, internal documents, precedents, document-management systems, citations, review processes and firm-specific knowledge.

That surrounding workflow is already large enough to support more than half a billion dollars in combined annualized revenue across these two companies.

Company	Recent annualized revenue	Recent scale
Harvey	$350M+	Revenue up more than 80% from the start of the year
Legora	~$150M	Around 1,500 customers and 100,000+ legal users
Combined	$500M+	Two large businesses built above outside foundation models

Can a GPT wrapper actually be profitable?

Yes, some model-dependent AI applications are already profitable, although the economics vary wildly from one product to another.

Gamma gives us one of the cleanest examples. Stripe says the AI presentation company passed $100 million ARR after reaching 70 million users, while Gamma says it had already been profitable for two years. It reached that scale with roughly 50 employees and only $23 million of initial funding.

Chatbase has also grown without outside funding. Stripe says the company crossed $10 million ARR with 26 employees roughly three years after launch, while Supabase reported more than 8,000 paying customers.

Cursor shows the other side of the economics. The Information reported a gross margin of negative 23% for one quarter earlier this year, largely because Cursor was spending so much on the models powering its coding agents. The publication later reported that gross margins had moved back into positive territory.

Coding agents can be exceptionally expensive. They may read large repositories, generate code, inspect the output, run tools, retry failed approaches and continue working for long periods. A presentation generator or customer-service agent may consume far less frontier-model compute relative to the subscription price.

Falling inference prices help. Stanford's AI Index found that the cost of using a model with roughly GPT-3.5-level benchmark performance fell by more than 99% within two years. Yet AI products are simultaneously asking models to do much larger jobs, so cheaper tokens do not automatically turn into higher margins.

Company	Revenue scale	What we know about profitability
Gamma	$100M+ ARR	Profitable for two years when it disclosed the milestone
Chatbase	$10M+ ARR	Bootstrapped
Cursor	$4B+ annualized	Had a quarter with negative gross margin before recovering

Can a small GPT wrapper still reach $10 million in revenue?

Yes, a relatively small GPT-wrapper business can still reach eight figures if it finds a narrow job customers repeatedly pay to automate.

Chatbase is probably the clearest current example. Founder Yasser Elsaid launched it during the first wave of LLM applications in 2023. The original product let users upload their own information and quickly create a chatbot around it.

That was close to the classic wrapper formula.

Three years later, Stripe reports more than $10 million ARR. The company has only 26 employees, remains bootstrapped and has expanded toward customer-service agents that connect business information with real support workflows. Supabase reported more than 8,000 paying customers earlier this year.

The business therefore reached eight figures without training a frontier model, raising a giant venture round or trying to compete with OpenAI on raw intelligence.

Jenni AI offers another, less independently documented example. The academic-writing company has publicly put its revenue around $12 million ARR after focusing on research, citations and student workflows rather than staying a generic writing generator. We would give the Chatbase number more weight because Stripe and Supabase provide stronger third-party confirmation.

Neither business changes the difficulty of launching another generic AI tool today. They do show that the ceiling for a focused, small-team wrapper remains much higher than many people assume.

What happened to Jasper and Copy.ai after ChatGPT attacked AI writing?

Jasper shows how quickly a horizontal GPT wrapper can get squeezed, while Jasper and Copy.ai also show the obvious escape route: move deeper into a business workflow.

Jasper was already around $90 million ARR when its growth ran into ChatGPT. The Information reported that the company cut its annual recurring revenue projections by at least 30%, reduced staff and shifted away from a broad consumer-plus-business strategy.

The underlying problem was easy to see. General-purpose AI had become good enough that customers no longer needed a separate product simply to generate competent marketing copy.

Jasper subsequently pushed much harder into enterprise marketing, with brand controls, company knowledge and workflow automation. The company has said its enterprise ARR later tripled over a one-year period, although we do not have a strong enough recent absolute revenue figure to compare Jasper directly with Harvey or Gamma today.

Copy.ai made a similar move. Its original copy generator became increasingly hard to distinguish from what users could produce inside ChatGPT. The company repositioned around go-to-market workflows and reported 480% revenue growth during 2024, followed by several months of more than 20% monthly ARR expansion. Again, recent absolute revenue is much less transparent than the percentage growth.

Jasper and Copy.ai are not good proof of today's largest wrapper businesses. Their histories are more useful as a warning: generic generation gets commoditized quickly, and surviving usually requires moving closer to the customer's actual workflow.

Do better AI models help GPT wrappers or eventually kill them?

Better AI models are helping the strongest application companies today because each model improvement makes their existing workflow more capable.

Imagine a generic rewriting app. When Claude gets much better at rewriting, Claude itself becomes a stronger alternative to that app. The wrapper receives a better model and a more dangerous competitor at exactly the same time.

Harvey experiences the upgrade differently. A better reasoning model can immediately work across the legal documents, permissions, integrations and firm knowledge Harvey already organizes. Customers get a stronger legal product without rebuilding their entire workflow.

Lovable gets another advantage from better coding models because increasingly complex applications become possible through the same product. Cursor can delegate larger programming tasks. Gamma can generate better visual material.

This helps explain why model commoditization has produced such uneven outcomes. Companies selling access to intelligence lose differentiation as intelligence improves. Companies applying that intelligence inside a valuable workflow can sometimes improve every time their suppliers release a better model.

The strongest application businesses have surprisingly little reason to hope foundation-model progress slows down. They want the models to become cheaper, smarter and easier to switch between.

Are OpenAI, Anthropic and Google now coming directly for GPT wrappers?

Yes, the foundation-model companies are currently moving deeper into application software, and that creates a serious threat for wrappers that still live too close to the model.

Coding makes the pressure obvious. OpenAI says Codex is now used by more than four million people per week and by companies including Cisco, Datadog, Dell and Nvidia. Anthropic has turned Claude Code into a major developer product. Cursor now competes directly with two of its most important model suppliers.

Legal AI is starting to look similar. Anthropic recently expanded Claude for legal work with more than 20 connectors and a dozen specialized plugins covering legal systems and workflows. Google has just introduced Gemini Enterprise for Legal, which connects its models to legal data, firm systems and agentic workflows.

Those releases move the foundation labs much closer to what Harvey and Legora sell.

The outcome is still open. Google's new legal product also includes Legora and Harvey within its partner ecosystem, which shows that the labs are willing to supply and compete with application companies at the same time.

This pressure should keep rising. Every major AI lab wants more of the software layer above its models, especially in coding and professional work. A wrapper needs enough product depth that the customer cannot replace it simply by opening a new feature inside ChatGPT, Claude or Gemini.

Why would anyone pay for a GPT wrapper instead of just using ChatGPT?

People pay GPT-wrapper companies because repeatedly completing an entire job inside ChatGPT can still be much more cumbersome than using software built around that job.

A lawyer can already ask ChatGPT or Claude to review a contract. A large law firm still needs the correct client documents, access permissions, prior matters, approved templates, citations and internal knowledge to appear automatically around that request. Harvey and Legora handle much of that machinery.

Developers can paste code into Claude. Cursor gives the model continuous access to the repository, applies edits across files, runs tools and keeps the model inside the development loop.

A founder can ask ChatGPT to generate application code. Lovable turns the request into a project that can be iterated on and deployed without requiring the founder to manually assemble everything the model produced.

Gamma follows the same pattern for presentations. ChatGPT can produce an outline and slide text, while Gamma gives the user an editable visual artifact that can be shared and continued inside the same product.

The amount of work removed around the model often determines how much customers will pay. Raw generation has become cheap. A completed workflow can still be worth hundreds or thousands of dollars per user each year.

What do the successful GPT wrappers actually own?

Successful GPT wrappers increasingly own the workflow, customer context and distribution around the model, which gives them something valuable even when the underlying AI supplier changes.

Cursor controls where developers work with AI and how their repositories enter the model context. Harvey and Legora connect models to legal matters, institutional knowledge and professional systems. Lovable owns much of the path between an idea and a deployed application. Gamma owns the environment where generated presentations and documents remain editable.

They are also becoming more comfortable using several models at once.

Harvey has openly described its multi-model approach across OpenAI, Anthropic and Google. Cursor lets users choose among several outside model families and its own models. Lovable offers outside frontier models while developing internal models. Many other AI applications now route different jobs toward different providers.

This changes the relationship with the labs. OpenAI or Anthropic can remain extremely important suppliers without being the whole product.

The strongest companies are building enough around the model that changing suppliers becomes a technical decision rather than a complete product reset.

Which GPT wrappers look most vulnerable today?

Generic writing, summarization, document chat and other thin GPT wrappers look increasingly vulnerable today because general AI assistants already handle those jobs well.

A product becomes especially exposed when a customer can move to ChatGPT, Claude or Gemini without losing much history, configuration or workflow.

Basic rewriting fits that description. Generic brainstorming does too. Simple PDF chat has become much easier to reproduce inside general assistants as context windows and file handling have improved. Basic email generation and undifferentiated chatbot builders face similar pressure.

The picture changes when customer usage leaves something behind: internal data connections, project history, workflows, organization-specific instructions, permissions, integrations or collaboration between employees.

Jasper's early slowdown showed what happens when a foundation-model interface catches up with a wrapper's main feature. Harvey, Cursor, Lovable and Gamma have taken the opposite route and added more product between the model and the final outcome.

That gap around the model is currently where much of the application-layer value is accumulating.

So, are any GPT wrappers still making a lot of money?

Yes, GPT wrappers are still making a lot of money, but the companies making the most have usually grown far beyond the thin-wrapper model that became popular after ChatGPT launched.

We can now point to multiple businesses at serious scale rather than one freak example. Cursor has reached billions of dollars in annualized revenue. Lovable is above half a billion. Harvey and Legora together are around half a billion. Gamma has crossed $100 million ARR while reporting profitability. Chatbase shows that even a bootstrapped team can still get beyond $10 million.

At the same time, the original criticism of GPT wrappers has aged pretty well. A startup that mainly takes a user's prompt, sends it to somebody else's model and presents the answer in a slightly better interface has very little protection these days. ChatGPT, Claude and Gemini keep absorbing those obvious use cases.

The strongest businesses have moved much closer to the work customers are trying to finish. They decide what context the AI sees, connect it to the right systems, let it take actions, check its work and carry the result into the next step.

Revenue quality remains an important caveat. Annualized run rates can exaggerate how predictable fast-growing AI revenue really is, and Cursor's earlier negative gross margin shows that enormous sales can coexist with expensive model bills. Platform risk is growing as well because OpenAI, Anthropic and Google increasingly build their own coding, legal and professional applications.

Still, the current evidence leaves little room for the broad claim that GPT wrappers are dead.

The thin ones are having a much harder time. The best ones have turned into full software companies, and some of those companies are currently making extraordinary amounts of money.