Company

What is an AI wrapper? Meaning, examples and the moat debate

An AI wrapper is a software product whose core intelligence comes from a model built by another company and reached through that company's API. The wrapper supplies the interface, the instructions sent to the model, the integrations and the billing, while a provider such as OpenAI, Anthropic or Google does the actual text, code or image generation. The label is informal and usually critical: it suggests a product that a rival, or the model provider itself, could copy quickly.

In practice the term covers everything from weekend side projects to some of the most valuable private software companies of recent years. Company facts below are as of October 2026.

What an AI wrapper means in practice

Technically, a wrapper is simple. The application takes a user's input, adds its own instructions and context, and sends the bundle to a model provider's API, for example Anthropic's Messages API. The model returns a response, which the application formats and shows to the user. The wrapper company does not train or host the model; it rents access to it.

That access is metered. Providers price API usage per token, the small chunks of text a model reads and writes, with separate rates for input and output, as Anthropic's API pricing page shows. Every user action therefore carries a direct cost that traditional software does not have.

"GPT wrapper" means the same thing, with OpenAI's GPT models underneath. In both forms the criticism is identical: if most of the value comes from the model, a competitor with the same API access could rebuild the product.

The wrapper spectrum: from thin layer to own model

"Wrapper" is less a category than a position on a scale, and many companies move along it over time.

1. Thin wrapper. A prompt template and an interface around one model, such as a basic tool that summarizes uploaded documents. Cheap to build and just as cheap to copy.
2. Workflow product. Model calls sit inside a multi-step process with integrations, permissions, review steps and saved history. In its 2023 essay Generative AI's Act Two, Sequoia described this shift as treating foundation models as one part of a fuller solution rather than the whole product.
3. Vertical AI. A product built for one industry, with domain-specific evaluations, data connectors and compliance work, as in legal, healthcare or finance.
4. Model-owning application. The company trains or post-trains its own models for core tasks and uses outside models selectively. Further still are research-first companies that train frontier models from scratch, sometimes called AI neolabs.

Most mature products route requests between several models, their own and other labs', depending on cost and task, so "wrapper" is a matter of degree.

AI wrapper examples

Cursor: from routing models to training them

Cursor, the AI code editor made by Anysphere, gives developers access to frontier models from outside labs. In a July 2025 pricing post, the company said it combines its own custom models with models from providers including OpenAI, Anthropic and Google. In October 2025 it released Composer, an in-house agent model for coding that it trained with reinforcement learning on real-world software engineering tasks.

In November 2025 Cursor announced a $2.3 billion Series D at a $29.3 billion post-money valuation and said it had passed $1 billion in annualized revenue. In April 2026 it partnered with SpaceX to train models on SpaceXAI's Colossus infrastructure, saying compute had been its bottleneck. On August 14, 2026 it said it had been acquired by SpaceX. It is a clear case of a company moving from the wrapper end of the scale toward the model-owning end.

Harvey: vertical AI that went multi-model

Harvey sells AI tools to law firms and in-house legal teams. In May 2025 it said it was supplementing its existing OpenAI models with models from Anthropic and Google, arguing that no single model was best at every legal task. In August 2026 it previewed Harvey Tenet, its first post-trained open-weight model, built on a Kimi K3 base together with Fireworks, and pointed to the lower per-token prices of open-weight models.

In September 2026 Harvey raised $550 million at a $15.5 billion valuation and said 80% of Am Law 100 firms use its product.

Windsurf: what platform risk looks like

Windsurf, another AI coding editor, shows the downside of depending on a supplier. In June 2025 Windsurf said Anthropic had cut nearly all of its first-party capacity for Claude 3.7 Sonnet and Claude 3.5 Sonnet with less than five days' notice, TechCrunch reported, at a time when OpenAI was reported to be buying the company. Anthropic said it was prioritizing capacity for sustainable partnerships.

The OpenAI deal then fell apart. In July 2025 Google DeepMind hired Windsurf's chief executive, a co-founder and some of its top researchers, and Google took a nonexclusive license to some Windsurf technology, TechCrunch reported. Citing Bloomberg, it put the price at $2.4 billion. The arrangement is known as a reverse acquihire. Days later Cognition agreed to acquire the remaining business, saying Windsurf had $82 million in annual recurring revenue.

The main risks of an AI wrapper business

Platform risk

A wrapper's key supplier controls access, prices, rate limits and the model lifecycle. Anthropic's deprecation policy, as of October 2026, promises at least 60 days' notice before a publicly released model is retired, after which requests to it fail. The Windsurf episode shows that access itself can change at short notice.

Inference costs and margins

Because every request costs money, AI applications tend to carry lower gross margins than classic software. In January 2023 Andreessen Horowitz estimated that generative AI app companies spent roughly 20–40% of revenue on inference and per-customer fine-tuning, with gross margins more often in the 50–60% range than near 90%. Bessemer's State of AI 2025 report put gross margins for its fastest-growing group of AI companies, which it calls Supernovas, at about 25% and often negative, against about 60% for a steadier group it calls Shooting Stars.

Costs also vary sharply between users. Cursor said in its July 2025 post that its hardest requests cost an order of magnitude more than simple ones, which is why it moved from per-request limits to usage-based pricing. Common cost levers include routing simple tasks to cheaper models, caching repeated context and batching work that is not urgent. As of October 2026, Anthropic lists a 50% discount for its Batch API.

The model provider becomes a competitor

Model labs build applications too. Anthropic introduced Claude Code, an agentic coding tool, as a research preview in February 2025 and made it generally available in May 2025, along with beta extensions for VS Code and JetBrains. That put a model supplier in the same market as the coding editors that buy its models.

What investors look for instead of a thin wrapper

Using third-party models does not by itself rule a company out: Cursor and Harvey both raised large rounds while still using models from outside labs. The question is what the company owns that the model provider does not:

• Proprietary data. Customer documents, outcome data and task-specific evaluations that a new entrant cannot easily obtain. Sequoia's 2023 essay added a caution: data moats had proved weaker than its authors predicted, except in fields where specialized data is hard to get.
• Workflow lock-in. Integration with systems of record, approval chains and team processes, so switching means retraining staff and rebuilding pipelines. Sequoia named workflows and user networks as the more durable advantages, and Bessemer has argued that context and memory may become new moats.
• Distribution. Existing customer relationships, a sales channel into a regulated industry or a brand users ask for by name.
• Model independence. The ability to swap providers or run its own models, as Harvey and Cursor have done, which limits supplier risk and can improve margins.

Recent rounds for AI application companies are tracked in our Funding section.

Is an AI wrapper a bad business?

Not by definition. Building on someone else's model lets a small team ship quickly and test demand before paying for its own compute. The weakness shows when a product never moves beyond the thin end of the scale. Founders can check their position with a few questions:

• What happens if the model provider ships this feature for free?
• What happens to margins if the main model's price doubles, or if access is cut for a week?
• What does the product know about each customer after six months that a new rival would not?
• Could the product switch to another model provider within days?

Early compute bills are one reason teams apply for startup cloud credits. For more on how AI application companies make money, see our AI Business coverage.

Frequently asked questions

What does GPT wrapper mean?

A GPT wrapper is an AI wrapper built on OpenAI's GPT models: an app that sends user requests to OpenAI's API and shows the answers in its own interface. The term is often used dismissively.

Is ChatGPT an AI wrapper?

Not in the usual sense. ChatGPT is OpenAI's own application running on OpenAI's own models, while the wrapper label describes a product one company builds on another company's model.

Is Cursor an AI wrapper?

Cursor still offers models from outside labs, but since October 2025 it has also shipped its own Composer coding models. In August 2026 it said it had been acquired by SpaceX, which puts it near the model-owning end of the spectrum.

Are AI wrappers profitable?

Some are, but paying for inference on every request makes margins thinner than in classic software. Bessemer's 2025 report put gross margins for the fastest-growing AI companies it studied at about 25%, often negative.

How do you build an AI wrapper?

The basic version needs an API key from a model provider, a prompt that frames the task and an interface that passes user input to the API and displays the response. The harder part is building data, workflows and distribution that a competitor with the same API cannot copy.