Almost every AI tool you’ve used — a writing assistant, an image generator, a customer service chatbot — is built on top of something called a foundation model. Understanding what that term means helps explain why so many seemingly different AI products are, under the hood, more related than they appear.
What Is a Foundation Model?
A foundation model is a large AI model trained on a broad, massive dataset that can then be adapted to many different downstream tasks, rather than being built to do just one thing. Instead of training a separate model from scratch for every use case — one for translation, one for summarization, one for customer support — companies build one large, general-purpose model, then adapt or build on top of it for specific applications.
GPT, Claude, Gemini, and Llama are all examples of foundation models. They’re trained once, at enormous cost, on broad general knowledge and language patterns — and then that single trained model becomes the foundation for countless different products and features.
Why “Foundation”?
The term reflects the role these models play architecturally: they’re the base layer that other things get built on top of. A company building an AI writing tool, for example, doesn’t need to train its own language model from scratch — it can build its product on top of an existing foundation model (through an API), adding its own interface, specialized prompts, and business logic on top.
This is why so many AI tools that look completely different on the surface — a legal research assistant, a recipe generator, a coding helper — may actually be powered by the exact same underlying foundation model, just adapted and prompted differently for their specific purpose.
How Foundation Models Get Adapted
Companies building on top of a foundation model typically use one or more of these approaches:
- Prompting: giving the model detailed instructions and context at the time of use, without changing the model itself — the simplest and most common approach.
- Fine-tuning: further training the foundation model on a smaller, specialized dataset to make it better at a specific task or domain.
- Retrieval-augmented generation (RAG): connecting the foundation model to an external knowledge source, like a company’s own documents, so it can reference current, specific information it wasn’t originally trained on.
Why This Matters for Evaluating AI Software
When comparing AI tools, it’s worth knowing which foundation model a product is actually built on, since that often explains a lot about its strengths, weaknesses, and pricing. Many AI startups are, in practice, building a specific workflow or interface on top of the same handful of foundation models from OpenAI, Anthropic, Google, or Meta — the real differentiation often comes from the data, workflow, and user experience layered on top, not the underlying model itself.
The Bottom Line
A foundation model is the broad, general-purpose AI model that sits underneath a huge share of today’s AI products, adapted through prompting, fine-tuning, or retrieval rather than being built from scratch each time. Recognizing this layered structure makes it much easier to understand what you’re actually using — and paying for — when you pick up a new AI tool.