glossary
What Is a Foundation Model in AI? Definition
A foundation model is a large, broadly trained AI model general enough to be adapted to many tasks through prompting or fine-tuning, rather than built for one narrow job.
5 min read · Updated Jul 1, 2026
A foundation model is a large AI model trained on broad, diverse data at massive scale so it can be adapted — through prompting or fine-tuning — to many different downstream tasks, rather than being built for just one job. GPT-5.2, Claude Opus 4.6, and Gemini 3.x are all foundation models; ChatGPT, Claude.ai, and the Gemini app are the consumer products built on top of them.
Why they’re called “foundation” models
The term was popularized around 2021 by Stanford’s Center for Research on Foundation Models, which used it to describe a new class of large, general-purpose pretrained models that serve as a shared base — a “foundation” — for many downstream applications, instead of each application needing a model trained from scratch. The name has since become the standard industry term for this category.
Foundation model vs. LLM vs. fine-tuned model
These three terms describe layers, not competing categories:
| Term | What it means |
|---|---|
| Foundation model | A large, broadly trained base model, general enough for many tasks |
| LLM (large language model) | A foundation model specialized for text — the most common type consumers interact with |
| Fine-tuned model | A foundation model further trained on a narrower dataset to specialize it for one task or tone |
In short: an LLM is usually a type of foundation model, and a fine-tuned model is a foundation model that’s been customized. See our AI model overview for how these fit into the wider picture, and our fine-tuning explainer for how the customization step actually works.
How a foundation model is actually trained
Training a foundation model starts with pretraining — the model is shown enormous amounts of text (and, for multimodal models, images, audio, or video) and learns to predict what comes next, gradually absorbing patterns of grammar, facts, reasoning, and style along the way. This stage requires massive compute and is why only a handful of well-funded labs — OpenAI, Anthropic, Google DeepMind, xAI, and a few others including DeepSeek — build frontier foundation models from scratch. After pretraining, labs run additional alignment passes (teaching the model to follow instructions and avoid harmful output) before releasing the finished foundation model into a product.
Affiliate link for Perspective AI — we may earn a commission at no extra cost to you.
Examples of foundation models powering 2026 subscriptions
| Foundation model family | Developer | Consumer product |
|---|---|---|
| GPT-5.2 family (GPT-5, GPT-5.4, o3) | OpenAI | ChatGPT |
| Claude (Opus 4.6, Sonnet 4.6, Haiku) | Anthropic | Claude.ai |
| Gemini 3.x | Google DeepMind | Gemini app, Google Workspace |
| Grok 4 family | xAI | Grok |
| DeepSeek models | DeepSeek | Available directly and through aggregators |
Because each product is a thin layer over a different foundation model, comparing subscriptions really means comparing the foundation models underneath — see our pricing hub for a breakdown of each provider’s plans. Higher-priced tiers within the same provider often unlock a stronger version of the same foundation model family rather than a completely different one, which is worth checking before assuming a pricier plan buys you a fundamentally different model.
How companies build on top of foundation models
Rather than training a new model from scratch, developers and companies typically adapt an existing foundation model using one of a few lighter-weight techniques:
- Prompting — giving the model instructions and context at the moment of use, with no retraining involved.
- System prompts — persistent background instructions that shape the model’s behavior across a whole conversation or product.
- Retrieval-augmented generation (RAG) — feeding the model relevant external documents at request time so it can answer using up-to-date or private information.
- Fine-tuning — additional training on a narrower dataset to bake in a specific tone, format, or domain expertise; the most resource-intensive option.
- Agents — wrapping the foundation model in planning and tool-use logic so it can take multi-step actions, covered in our AI agent explainer.
Platforms like Perspective AI take a different approach entirely: instead of building on a single foundation model, they license access to many — GPT-5.2, Claude, Gemini, DeepSeek, and more — and let you pick the best one for each task from a single subscription.
Why this matters when picking a subscription
- The foundation model determines the ceiling. No amount of clever app design fixes a weaker underlying model.
- Different foundation models have different strengths — Claude for writing quality, GPT-5.2 for feature breadth, Gemini for context length and Workspace integration.
- Subscription price often tracks foundation model access, not just app features — higher tiers usually unlock stronger or additional foundation models, not just more of the same one.
- You’re not stuck with one — comparing options across providers, or using an aggregator, lets you match the foundation model to the task instead of the task to whichever model you already pay for.
Because pretraining a frontier foundation model from scratch is so expensive, most of the differentiation you’ll notice between providers over time comes from what’s built on top — fine-tuning, system prompts, agent tooling, and app-level features — rather than an entirely new base model appearing every month.
Going deeper
Frequently asked questions
What is a foundation model, in plain terms?
A foundation model is a large AI model trained on broad, diverse data at massive scale so it can be adapted to many different tasks — through prompting or fine-tuning — instead of being built to do just one thing.
Why are they called 'foundation' models?
The term was popularized around 2021 by Stanford's Center for Research on Foundation Models to describe large pretrained models that serve as a general-purpose base other tools, products, and fine-tuned models are built on top of.
Is GPT-5.2 a foundation model?
Yes. GPT-5.2 is a foundation model developed by OpenAI, and ChatGPT is the consumer product built on top of it, alongside GPT-5.4, o3, and other models in the same family.
What's the difference between a foundation model and an LLM?
A large language model (LLM) is a foundation model specialized for text. 'Foundation model' is the broader category, which also includes multimodal and image-focused models that aren't purely language-based.
What's the difference between a foundation model and a fine-tuned model?
A foundation model is the broad, general-purpose base. A fine-tuned model is that same base after additional training on a narrower dataset to specialize it for a specific task, tone, or domain.
Which companies build foundation models?
OpenAI (GPT-5.2 family), Anthropic (Claude Opus/Sonnet/Haiku), Google DeepMind (Gemini 3.x), xAI (Grok 4), and DeepSeek all develop their own foundation models, which power the consumer subscriptions built on top of them.
Do I ever interact with a foundation model directly?
Rarely as a consumer. You typically interact with an app — ChatGPT, Claude.ai, Gemini — that wraps the foundation model in a chat interface, memory, and safety layers. Developers can access foundation models more directly through an API.