An AI model is the part of an AI that learned from examples during training and then computes an output from your input. Technically it is a huge collection of numbers, the parameters, together with the blueprint for how to calculate with them. The model is not the same as the app: ChatGPT is the app, GPT-6 is the model behind it and OpenAI is the vendor.
For me, the difference matters in practice every day. My fourteen AI employees run on Claude Team, some also on open models or through the API. An employee's role stays the same; the model underneath can change. Once you understand that, you also understand why the same question gets different answers in two tools. All AI basics are on my AI Basics page.
What an AI model is
In its machine learning glossary, Google describes a model in general as any mathematical construct that processes input data and returns output. Put differently: a model is the set of parameters and structure a system needs to make predictions.
According to Google, parameters are the values a model learns during training, mainly so-called weights. They determine how strongly each part of the input counts in the calculation. Large language models have many billions of them. In practice, two things happen:
- Training: the model sees a huge number of examples, and its parameters are adjusted step by step until its predictions are good. This happens at the vendor, once per model version, and takes a lot of computing power.
- Use: you send a question, and the finished model computes an answer with the learned parameters. It doesn't learn anything new in the process. Experts call this inference.
What kind of output a model produces depends on its training: a language model generates text, an image model images, other models sort emails or recognise faces. The models that create new content, I explain in Generative AI Explained, with Examples.
Model, app and vendor: three different things
Most mix-ups happen because three things have similar names.
At Google, the model and the app are both called Gemini, which makes it particularly confusing. What matters: the app is the interface with the chat window, file storage and settings. The model is the computing core that generates the answers. The same app can offer different models, and the same model can sit in many apps, including programs from other companies that plug it in through the API. The API is the interface a program uses to talk to a model directly, without a chat window. How that works, I explain in What Is an API?.
What model names tell you
Model names usually follow a pattern of family, tier and version. In Claude Opus 5.5, Claude is the family, Opus the tier and 5.5 the version. Anthropic describes the tiers like this: Opus for long tasks in coding and knowledge work where the AI takes many steps on its own, Sonnet as the best combination of speed and intelligence, Haiku for high volumes of fast requests, Fable for especially demanding reasoning. Which Claude model suits what is in Claude Models: Opus, Sonnet, Haiku, Fable.
Every model also has its own knowledge cutoff. Anthropic gives June 2026 for its current models. OpenAI lists 30 April 2026 for GPT-6 Astra and GPT-6.1 Sol, and 18 May 2026 for GPT-6 Luna. If you wonder why a tool doesn't know about a recent piece of news, that is often the reason.
With open models, the name often includes the size, such as "8B" or "70B". The B stands for billion parameters. Larger models can often do more, but they also need more memory and a more powerful computer.
Open and closed models
- Closed models you can only use through the vendor's app or API. The parameters stay with the vendor. Claude and the GPT models in ChatGPT are among them. For Claude, Anthropic doesn't even give the number of parameters in its model overview.
- Open models, often called open-weight models, you can download and run on your own hardware. Weights here means the learned parameters. Whether you may use an open model commercially depends on its licence.
Which open models exist is in Open Source AI Models 2026 Compared; what really works with them on your own hardware is in Local AI: What Actually Works in 2026. I explain the licences in Open Weight vs Open Source: Licenses.
An example: the same task, three routes
A made-up example: an electrical contractor wants quote texts written from bullet points. There are three routes. First, a vendor's app, where the owner picks a model from the list and starts typing. Second, a program that always addresses the same model with the same instruction through the API and writes the text straight into the quoting software. Third, an open model on the business's own computer in the office; then the bullet points never leave the building, but someone has to run the machine.
The task is the same three times; the model and the route to it are not. It's the same with my AI employees: the role stays, and I can switch the model underneath.
What to watch out for
- Ask which model is inside a tool. A program with "AI" in its name may have its own model or use one from a big vendor through the API. Quality and the path your data takes depend on it.
- Check again after a model change. A new model may write differently, even with the same instruction.
- Expect models to be retired. For each current model, Anthropic gives an earliest date on which it will be switched off; for Opus 5.5 that is 22 September 2027 at the earliest.
- Know the knowledge cutoff. For current topics the model needs a web search or your material.
- With open models, read the licence before you use them for client work.
Frequently asked questions
Is ChatGPT an AI model?
No, ChatGPT is an app. According to OpenAI, the model behind it has been GPT-6 since October 2026. In the same way, Claude is Anthropic's app, in which models such as Claude Opus 5.5 or Sonnet 5.5 do the work.
What is the difference between an AI model and an AI agent?
The model generates an answer to an input. An agent uses a model as its engine, gets tools such as web search or files, and takes several steps in a row until a task is done. More in What Is an AI Agent? Explained Simply.
What do numbers like 8B or 70B in a model name mean?
They indicate size: 8B means 8 billion parameters, 70B means 70 billion. Larger models can often do more but need considerably more memory. For an ordinary office computer, smaller models are the better fit.
What happens when a vendor retires a model?
Your work then continues with the successor, and results may change. Anthropic publishes an earliest retirement date for each model. If you have set up a tool permanently, plan a test with the new model in good time.
Does this have anything to do with AI models on Instagram?
No. In English, "AI model" sometimes also means figures on Instagram whose pictures are generated entirely by an AI. This post is about the model as software, the computing core of an AI.
Where to go from here
How a model learns in the first place, I explain in my post on machine learning. Which models create new content is in Generative AI Explained, with Examples. The next step on the learning path is the inside of a language model: How Do LLMs Work? A Plain-English Guide. How others choose their models is what we discuss in my German-speaking community Claude Practitioners.
