What does the term "AI model" mean?

You will often hear phrases like "the latest model," "switch models" or "a more powerful model." An AI model is the trained system at the core of an AI product, the part that actually does the thinking, writing or recognising.

A helpful way to understand it is to separate the model from the app. ChatGPT, Claude and Gemini are apps or products, the websites and mobile apps you use. Inside each one is a model, or several models, that generate the answers. The app is like a car's body and dashboard; the model is the engine. The same company can put a new engine in the same car, which is why the app stays familiar while its abilities improve.

How is a model made? Engineers start with an architecture, a design for a neural network. They train it on huge amounts of data, adjusting billions of internal numbers called parameters until the model gets good at its task. The result, a file containing all those learned numbers, is the model. It can then be used again and again to answer new questions.

There are many kinds of AI models:

Language models, which work with text, like those behind ChatGPT or Claude.
Image models, which create or recognise pictures.
Speech models, which turn speech into text or text into speech.
Embedding models, which turn text into numbers for search.
Multimodal models, which handle several types of input together.

Companies usually release families of models in different sizes. Smaller models are faster and cheaper but less capable. Larger models are more capable but slower and more expensive to run. Anthropic, for example, has Haiku, Sonnet and Opus models, and OpenAI and Google have their own ranges. In many apps, you can choose which model to use from a menu.

Model names include version numbers that change frequently, as companies release improved versions every few months. Newer is usually better, but for simple tasks a smaller, faster model is often perfectly fine.

Some models are closed, used only through the company's own products and APIs. Others are open-weight, meaning anyone can download and run them, such as Llama, Mistral, Qwen or DeepSeek.

So when someone says "the model," they simply mean the trained AI system doing the work behind the screen.