Why AI Wrappers Keep Dying
Last week Anthropic released Claude Opus 4.7. Two days later they launched Claude Design.
The two announcements should be read together. Because they aren’t just new products. They’re a strategic signal that every founder, investor, and leader building on top of large language models should take seriously.
Let me explain why.
What actually happened last week
Claude Opus 4.7 is currently the most capable generally available model on the market. It hits 87.6% on SWE-bench Verified (up from 80.8%), tops MCP-Atlas for multi-tool orchestration, and Anthropic themselves describe it as “more tasteful and creative” on professional tasks — producing higher-quality interfaces, slides, and docs than any previous model.
Two days later, on April 17, Anthropic launched Claude Design. A tool that lets anyone create prototypes, pitch decks, one-pagers, and UI mockups without a design background. Built on top of Opus 4.7.
The market reacted immediately. Figma’s stock dropped 7% the same day.
Three days before the launch, Mike Krieger, Anthropic’s Chief Product Officer, had quietly stepped down from Figma’s board. The market missed the signal.
This is the pattern I want to talk about.
A seminar on AI moats — and eight pillars
A year ago I sat in on a seminar about AI moats and new business models. Eight competitive advantages were discussed as the real defensible positions in the AI era:
- Data advantage — access to large, high-quality datasets that continuously improve performance
- Proprietary algorithms — unique solutions to niche problems that generic AI can’t crack
- Infrastructure — superior compute resources and scalable architecture
- AI talent — the ability to attract top expertise and build an AI culture
- Network effects — platforms that become more valuable as more users join
- Integration — seamless deployment into existing workflows
- IP protection — patents, trade secrets, and regulatory advantages
- Ecosystem — AI integrated across multiple products, creating compound value
Looking at that list today, it strikes me that most AI startups I follow can tick off maybe one or two of them. Often none.
The wrapper problem
There’s a term that’s been making the rounds in English-speaking tech circles for a while now: AI wrapper. It’s usually used as a put-down — a product that’s “just a wrapper” around GPT or Claude, a thin layer of UI and prompts sitting on top of someone else’s model. The critique used to feel a bit unfair. After all, plenty of great companies are “just wrappers” around databases, or “just wrappers” around payment rails.
But here’s why the wrapper critique is landing harder now: the wrapped product is actively hunting the wrappers.
And I say this as someone who loves Lovable. I use it frequently. It’s beautiful product craftsmanship and proof of how fast you can build something useful when you stand on the shoulders of a foundation model.
But here’s the uncomfortable thought: Lovable, Cursor, Replit, Harvey, Legora, and dozens of other AI startups are built on top of LLMs from Anthropic or OpenAI. Their product is essentially a layer of UX, workflow, and prompt engineering on top of someone else’s model.
Which of the eight pillars do they actually own?
- Integration? Maybe.
- Network effects? Weak, if any.
- Proprietary algorithms? Hardly — the brain itself comes from Anthropic.
- Data advantage? Sometimes, but usually built on user interactions that aren’t unique to them.
And now Anthropic ships Claude Design — a tool that does what Lovable, v0, and Figma do, baked directly into the same subscription developers are already paying for.
Get Danar Mustafa’s stories in your inbox
Join Medium for free to get updates from this writer.
The question isn’t whether Anthropic can build a Lovable clone. The question is when.
When the platform competes with its own developers
What’s happening here isn’t new in tech history. Amazon built AWS into the backbone of the internet and then started launching products that competed with its biggest customers. Apple did the same with the App Store. Whoever builds the platform always has incentives — often irresistible ones — to move up the value chain.
But the speed is different this time. A few numbers worth putting in context:
- OpenAI has 900 million weekly active users and crossed $25 billion in annualized revenue in February 2026
- Anthropic has gone from $1 billion in annualized revenue in January 2025 to over $30 billion in fifteen months
- Claude Code alone reached $2.5 billion in run-rate revenue in February 2026 and accounted for more than half of all enterprise spending at Anthropic
These companies don’t just have the models. They have the money, the talent, the distribution, and — perhaps most importantly — they have the user data that tells them exactly which applications are ripe to be absorbed.
When OpenAI sees millions of prompts about legal contracts, they know there’s a Harvey-shaped market ready to take. When Anthropic sees how developers use Claude to build UI, they know there’s a Lovable-shaped market ready to take. They don’t have to guess which markets work. Their users tell them in real time.
The result? OpenAI is pushing into healthcare and education. Anthropic is positioning in law, finance, and now design. They’re no longer just API providers. They’re building products.
What does this mean for you as a leader?
If your business model looks like this — “we take a foundation model and package it nicely for industry X” — it’s time to be uncomfortably honest with yourself.
It doesn’t mean you’re doomed. But it does mean you have to be clear about where your moat lives. If the answer is “we have better UX” or “we have better prompts,” that’s not a moat. It’s a runway.
The real defensible positions will be things the model companies can’t replicate by flexing their muscles:
- Unique, proprietary data that isn’t on the open web. Not “we’ve seen lots of user interactions,” but data that requires years of collection, partnerships, or physical presence to build.
- Deep vertical expertise that only comes from living in an industry for a long time. Generalist AI can cover 80% of a lawyer’s work. But the last 20% — the part that actually pays the bills — requires understanding how a specific jurisdiction works, how a specific client thinks, how a specific court decides.
- Distribution and trust in regulated industries. Hospitals, banks, and government agencies don’t switch vendors in a week. Hold a position there and it becomes your moat, regardless of which model powers the product behind the scenes.
- Physical systems and data from robotics, IoT, sensors, medical devices. The kind of thing you can’t scrape from a webpage.
What we’re all missing
Here’s what actually worries me most, and what I think few people are talking about: the speed.
Claude Opus 4.7 shipped two months after Opus 4.6, which shipped two months after 4.5. That’s a cadence that makes it nearly impossible to build a product on today’s model and assume your differentiator will hold into next quarter. What was impossible for a generic AI to do in February is default functionality in April.
Which means every time you build a feature that “does X better than raw GPT/Claude,” the clock is already ticking. The next model release might make your feature pointless.
That doesn’t mean stop building. It means build things that get better as the foundation models get better — not things that become redundant.
Three concrete tips for leaders and organizations
1. Do an honest moat analysis this week. Walk through the eight pillars and check off the ones you actually own. Not the ones you hope to own. The ones you have today. If you tick fewer than two, make that your most important strategic question for the rest of the year. Not “how do we launch the next feature” but “how do we survive the next model release.”
2. Invest in data only you can collect. The most valuable thing you can do right now isn’t building a prettier prompt wrapper. It’s building data loops that require your specific position in the market. Consented customer interactions. Partnerships with players sitting on unique information. Measurements from physical processes. Anything that makes your model better at serving your customer than a generalist AI ever can.
3. Choose to be complementary, not parallel. The companies that survive the platform owners’ encroachment are rarely the ones trying to do the same thing better. They’re the ones doing something the platform owners don’t want to do — because it requires deep regulation, specialist knowledge, or physical presence. If your product can be described in one sentence as “Claude Design but for X,” that sentence should scare you. If it can be described as “the thing Anthropic will never build because it requires ten years of partnership with Y,” you have something.
Finally
I’m not saying Lovable disappears tomorrow. I’m saying we’re in a period where the model companies have the motive, the means, and the momentum to move up the value chain. Their self-imposed limits on what they build will shift — not based on whether they can, but based on what’s most profitable.
If your business model rests on the assumption that Anthropic or OpenAI will stay a friendly API provider, you’re building on sand.
The good news is that this is also a moment of opportunity. The companies that build real moats now — in data, distribution, vertical depth — will be the ones standing three years from now. And most of them haven’t been built yet.
So the question I’ll leave you with: if Claude ships your product next month, what do you have left?
What do you think? Are we heading into a phase where only those with real moats survive, or is there room for a new generation of application companies built on top of the foundation models? Drop your thoughts in the comments.
