One name, three layers
When someone says "I use ChatGPT," they have named a product. The model inside it — the trained network doing the thinking — has a different name, and it gets swapped for a newer one every few months without anyone asking you. Most bad tool decisions start with those two things being confused.
Every choice you make is really three choices stacked.
The model. A very large file of numbers produced by an expensive training run. GPT models from OpenAI, Claude from Anthropic, Gemini from Google, Llama from Meta, Mistral from Mistral AI, Qwen from Alibaba, DeepSeek from DeepSeek. Each family ships in sizes: a large one that is slow and expensive, a medium one most people should use, a small one that is fast and cheap. Same brand, very different capability.
The product. Everything wrapped around the model: the chat window, file upload, web search, memory of past chats, the ability to run code, connections to your email or your code repository. This layer is where most of your daily experience actually lives.
The route. How you reach it. A website, a phone app, a plugin inside your code editor, a command in your terminal, a pay-per-use API, or a model file running on your own laptop.
The same model can feel like two different tools
Priya uploads a 40-page supplier contract to two different websites. Both say they run the same open model, same version. One gives her a useful clause-by-clause summary. The other misses half the document.
Nothing is wrong with the model. One product sent the whole contract. The other chopped it up, searched for what looked relevant, and sent three paragraphs. The model answered the question it was given. It was given a worse question.
This is the most useful habit to build before you compare anything. When a tool disappoints you, ask which layer failed:
- The model was not capable enough for the task.
- The product never gave the model what it needed.
- The route limited it — a free tier quietly serving a smaller model, an editor plugin that can only see the file you have open.
Most complaints that sound like "this AI is stupid" are the second or the third.
The map, briefly
Four product names cover most of what people use:
- ChatGPT (OpenAI) — the app most people met first. Web, mobile, desktop, plus an API.
- Claude (Anthropic) — chat app, an API, and a widely used coding tool that runs in a terminal.
- Gemini (Google) — chat app, and woven into Search, Docs, Gmail and Android on Google's own terms.
- Copilot (Microsoft and GitHub) — the clearest proof that product is not model. Copilot is a brand covering several different assistants, and over its life it has run models from more than one vendor. Asking "is Copilot good?" is like asking "is a restaurant good?" without asking who is cooking.
Then there are models you can download and keep: Llama, Mistral, Qwen, DeepSeek, Gemma. These have no product of their own. You supply the product — an app on your laptop, or a hosting company that serves them cheaply. Two lessons from now they get proper treatment.
Find out who you are talking to
Before you judge any tool, find the model name. It is usually in a dropdown at the top of the chat, or in settings, or at the bottom of the answer. Free tiers often start you on the strong model and move you to a smaller one after a few messages, and they do not always announce it clearly.
If you cannot find out which model is answering you, that is information too. It means you are evaluating a product, not a model, and your conclusions will not transfer when the company swaps the engine next quarter.
For the rest of this course, when a comparison is made, it will say which layer it is about. Almost every argument on the internet about which AI is best fails to.
Before you move on