What Is an AI Model? Model, Product, Agent Explained

An AI model is the frozen result of a training run: a set of numerical values tuned on huge amounts of data, able to predict what logically comes next given what it has been fed to read. It is not a programme written line by line with rules, it is the finished output of a statistical calculation. Three words get mixed up constantly in everyday language: the model, the product that exposes it, and the agent that uses it to act. Telling them apart avoids a lot of confusion in an AI project.

What is an AI model, in practical terms?

Technically, a model is a file of parameters: millions or billions of numerical values, tuned automatically during a training phase run over huge volumes of text, image or sound. Once that training run ends, the model is frozen. It does not keep learning from talking to you, it applies what it already learned.

That file does nothing on its own. Something has to sit around it to take a question in and hand a response back. That is exactly where the confusion starts: that piece of software is already something other than the model.

What is the difference between a model, a product and an agent?

Three distinct layers stack on top of each other, and everyday language blends them together constantly.

The model is the parameter file itself: a specific, numbered version of GPT, Gemini or Claude is a model. The product is the application that exposes it to the public: ChatGPT, Claude.ai and Gemini are products, each with an interface, a chat history, safety settings, and sometimes a choice of several models. The agent goes further: it uses a model to decide on an action, then carries it out (reading a file, calling a tool, writing a document), before starting again with whatever result comes back.

The same model can therefore sit behind three very different use cases: a plain conversation, a consumer product, or an agent working inside a folder of files with no constant human oversight.

What are a model's parameters?

Parameters are the numerical values a model adjusted during training, something like billions of dials turned gradually until the model's predictions became reliable. The more parameters a model holds, the more nuance it can, in theory, capture.

But bigger is not automatically better. Every extra parameter comes at a cost: in compute, which shows up as price per million tokens, and in latency, the wait before an answer arrives. Picking the right model size for a given task is exactly what the cost, latency and quality decision framework covers: a larger model is only the right call when the task justifies what it costs.

Why do publishers offer several sizes of the same model?

Nearly every major publisher now follows the same range logic: a fast, cheap model for repetitive tasks, a mid-tier model for everyday use, and a reasoning model, slower and pricier, kept for complex work. OpenAI documents its own model families this way, and Anthropic publishes the same kind of overview for its Claude models.

That range is not a marketing accident, it answers directly to the trade-off between cost, latency and quality. Choosing a model means choosing a use case first, not a product name.

What is the difference between an open and a closed model?

An open model publishes its parameters: anyone can download it and host it on their own servers. A closed model can only be used through its publisher's API, with no access to the underlying parameter file itself.

What does an AI model not do?

A model does not check a live database of facts for every question: it generates an answer from what it learned during training, without verifying a source in real time. When a business needs the model to know its own current documents, the real question to settle becomes RAG or fine-tuning, not which model to pick.

A model also does not know what it does not know. It does not return an error message when a piece of information is missing, it produces a plausible answer with exactly the same confidence as a correct one. That is the starting point for understanding and limiting AI hallucinations.

Finally, a model's knowledge stops at the date its training run ended. Unless it is wired to a search tool, it knows nothing of what happened after that date, including recent events in its own industry.

What does Gemini 3.7 Flash, released on 13 August 2026, actually show?

Gemini 3.7 Flash is a clear illustration of what a model looks like in 2026. Released on 13 August 2026, it accepts text, image, audio and video as input, with a context window of up to one million tokens. It is a single parameter file, able to handle several types of content, then exposed through different products and different agents depending on how it gets used.