OpenAI GPT Store Is Not the End of Thin-Wrapper GPT Startups

It's a wake up call

It doesn’t matter how much you tweak and improve a car with a small engine to make it faster; it will always be worse than a car with a big and powerful engine.

This is the dilemma that so-called thin-wrapper AI startups like Jasper and Askyourpdf — businesses built as an application layer on top of an LLM like GPT-4 — face in competing against model-building companies like OpenAI, Google, or Meta, which release better models every year and new features every month, all of which threaten to kill entire spaces where these GPT wrappers are trying to survive.

ChatGPT’s birth was the inflection point. How could Jasper, one of the first GPT wrappers and likely the most successful, compete now against a GPT-4-backed ChatGPT with just some fine-tuning on top that increases the price of the service 4x as a result? Three-quarters of Jasper’s customers pay $80/month whereas ChatGPT Plus is just $20/month.

Once OpenAI gets out its bigger engine (GPT-4.5 or GPT-5, likely much cheaper than $80/month), it won’t matter how good is Jasper’s fine-tuning of GPT-4. It will always be the smaller engine trying to compete against the big one. This is no different than the Bitter Lesson; your craft ingenuity is worthless against scale, data, and compute.

And that’s Jasper, the other GPT wrappers are much worse off.

A silver lining for adaptive GPT thin-wrappers

Common sense says Jasper is dead, just doesn’t know it yet.

But although the company is not going through its best moment, it’s not giving signs of being insolvent. I think there’s a simple explanation for why thin-wrapper AI startups that are already successful (in terms of user base or revenue) should not fear being killed by subsequent models or competitive features like GPTs, which OpenAI announced back in November, or the GPT Store it just launched: Narrow but deep often beats broad and shallow.

OpenAI is like the NYT (pun intended) in that it wants to attract the largest number of customers. That requires versatile products like ChatGPT, which is a proto-general purpose AI system. Likewise, the NYT has a section for everything but specializes in nothing. That allows GPT wrapper specialists to fill up gaps in niches OpenAI won’t bother covering (that’s also the reason why funnier models like Grok or steerable ones like Mixtral can thrive). Similarly, tech magazines like TechCrunch, Wired, or The Information narrow their focus to avoid competing against the NYT directly.

Generative AI is growing into a heterogeneous market that will allow different types of players to survive just fine. (This is also why no single AI model will rule the world, a topic I will write about shortly.)

This strategy of narrowing deeply, unfortunately, doesn’t apply to most startups that entered the generative AI landscape too late. The “AI for x” differentiation worked fine the first few dozen times we saw it, now it’s an overcrowded ecosystem waiting for the bubble to pop (or for the GPT store to absorb it).

In 2021, the generative AI industry was non-existent. Jasper, like OpenAI but on a different level, created a tiny new market it went on to dominate (that’s pretty much Peter Thiel’s advice to startups). That was possible for a wrapper startup back then but once the market expanded and reached a critical mass, succeeding with a similar offer became almost impossible. There are always exceptions like Askyourpdf but they’re rare.

Many of these late thin wrappers are built on top of ChatGPT directly (Jasper was built on top of GPT-3, which gave it an edge because it’s harder to use than ChatGPT). The “talk to your document” idea was brilliant until OpenAI updated GPT-4 in October 2023 to allow people to do this directly, including other possibilities like combining data analysis, DALL-E 3, and browsing in a single chat session. One simple feature promised to wipe out an entire nascent space. Not later, during DevDay, OpenAI announced both GPTs and the GPT Store, doubling down on its stomping of vertical wrappers.

Now that the GPT Store is live, some GPT startups have migrated there to find luck not competing against OpenAI but leveraging its developing ecosystem. The thin wrappers that the GPT Store has sentenced to choose between engrossing the store’s inventory or death would have died anyway crushed under the sheer pressure of market competition. The ones that managed to amass a substantial user base from early growth with a generic product (like Jasper), or as latecomers with specific services (like Askyourpdf), could pivot to a safer place afterward by further niching down (that’s what Jasper did).

It wasn’t only Jasper’s early presence that secured it a long life, but also its ability to rapidly capture a loyal user base who was happy with the startup’s service at that (high) price. When OpenAI threatened it, Jasper was narrowed down to marketing. Focusing on growth, user experience, and a well-defined task case were the right approaches that most latecomers didn’t replicate. They were not doomed due to OpenAI’s business moves, as many people have suggested, but because the market didn’t need them anymore and they didn’t know how to stay relevant. Thin wrappers like Askyourpdf, could’ve chosen Jasper’s path.

It was a risky bet but I don’t think any entrepreneur riding the generative AI wave on top of GPT — including Jasper founders or indie hackers iterating week after week — ever assumed they were safe behind a unique idea or an early success. OpenAI was always one feature away from making them redundant. Most knew what they were getting into and did it anyway to fight for the tiny possibility of somehow getting out of their fragile position before OpenAI decided to wipe them out. Early adopters like Jasper had a significantly higher chance of making it but they were never the only ones that could.

Now that the GPT Store is out building a wrapper isn’t a sensible business move anymore. Better make an appealing GPT, but know your chances are extremely tiny.

The burden of one more layer of abstraction

So, it must be clear by now that this article is not about all thin-wrapper AI startups but for those who dared risk knowing what they got into and rejected the appealing but tricky offer of being absorbed by OpenAI’s GPT Store.

Jasper is fine. Perhaps some other thin-wrappers are, too. The Jasper executive worried that OpenAI had, overnight, killed their $1 billion company. It didn’t happen. Narrowing down while keeping a neat UX works great because the value of removing, from the user, the burden of one more layer of abstraction can’t be overstated. Everyone who works on tech, especially software knows this. The median tech literacy is so painfully low that some days I find it hard to believe how the world keeps going given how technologified it is. I will never get over the fact that the word “Google” is constantly up at the top most searched queries in Google Search.

OpenAI knows this just fine. It knows most people just don’t care about its or Jasper’s tech stack. They don’t care that Jasper or Askyourpdf are literally API wrappers. They only care about one thing, and it has more to do with psychology than business: whether the product solves their problem.

But OpenAI’s focus is somewhere else. It won’t simplify its products too much just to gain some rapid growth in exchange for long-term instability, just like the NYT writes about everything but keeps a reasonably high degree of editorial quality across the board (or they’d risk becoming Twitter; more readers but inherently painful to use).

ChatGPT was a rare yet logical instance of this approach by OpenAI. It’s simpler than earlier models like GPT-3.5. Overplaying this hand would however damage both its long-term goals and its synergic relationship with developers. Also — and I can’t emphasize this enough — if you truly believe ChatGPT is easy to use by the standards of an average person, I urge you to ask your friends or family how to do X task that you might believe is intermediate in complexity with ChatGPT. It’s easier than GPT-3.5, but it is by no means a simple tool like, e.g. the iPhone. We are far from that delicate design and polished production (if anything, because we don’t fully understand yet what goes on inside LLMs).

OpenAI’s broad focus and ChatGPT’s intrinsic complexity explain why thin-wrappers can survive an ecosystem that should theoretically leave OpenAI as the sole player.

It also explains why open-source efforts can never beat incumbents. Google, Apple, or Microsoft can turn a generative AI idea into an easy-to-use feature readily deployed into products and devices, and integrated across an entire ecosystem of tools that their millions of clients already use daily (in some sense, open source always wins because it becomes the foundation of the products of the incumbents, but I’m not sure that’s what they’d call a victory).

It also explains how Lensa App grew so hilariously fast at the end of 2022 despite the quality of the technology being laughable in comparison to much better existing models (e.g. Stable Diffusion or Midjourney) that were just too far out of reach from a typical non-techie person. An additional layer of abstraction implies the loss of millions of potential users. That’s why Midjourney, which uses Discord (a business move that hardly makes any sense — they’ve been dominating the AI image scene not because of that but despite that), or StabilityAI, which creates open-source models, are fighting an uphill battle against OpenAI’s DALL-E 3 they are doomed to lose.

Jasper and the other early winners of the thin-GPT-wrapper startup gold rush live opposite to OpenAI on the easiness-of-use spectrum, even those that are merely ChatGPT wrappers. ChatGPT isn’t trivial to use. It’s too versatile; its untapped potential is too great, yet the path to extract the full value from it remains jagged. Most people I know personally don’t have a clue how to make it work for them for even the easiest tasks, like having a dialogue instead of having it generate three entire paragraphs for every prompt.

As Ethan Mollick has repeatedly advised, getting the most out of generative AI requires using it a lot to familiarize ourselves with its idiosyncrasy. Successful thin wrappers like Jasper and Askyourpdf do one thing very well besides narrowing down: They dumb it down.

They take out that burden of one (or a few) layers of abstraction from the user by making it super clear what their service does: You have a PDF document and want to know what’s in it without reading. Simple. If I want ChatGPT to make a marketing task for me, as Jasper does by default, I can either learn prompt engineering, search across a sea of marketing GPTs, or pay Jasper a bit more. Paying double the price for something so straightforward is plain theft for those of you who think it is straightforward. Those of you who don’t may even feel you’re the ones doing the stealing. It’s a bargain price if you don’t have any other way to access that knowledge or ability.

It all depends on perspective. And OpenAI shouldn’t be surprised: Why are people paying for GPT-4 to the point that OpenAI now predicts a staggering $1.6 billion in ARR when you can use GPT-4 for free in Microsoft Bing, which in contrast hasn’t managed to get a single percentage point of usage out of Google Search.

We think asking for money is the most frictional thing in the world but we often forget that applied knowledge can be worth, for the people who don’t have it, all the money in the world.

Most thin-wrappers will die anyway but not because of ChatGPT’s existence or the many features OpenAI releases that, one after another, threaten a bunch of thin-wrappers every time. The secret sauce for long-lasting success, and remaining impermeable to OpenAI’s moves, is to liberate their users from the burden of abstraction — if they put a little more effort into user care, intuitiveness, and usability than the competition, they are much more likely to stay on top, or at least not fall off.

Narrow down and dumb down.