Everyone nods along when someone says "ooh, new AI model". But what actually makes it tick?
Let's pull an AI chat apart and see which bit the model really is.
You've probably used all four today without noticing.
Here's what each one actually is, how you'd spot it at work, and some places you've already met it.
The window you type in. It handles the look and feel, your saved chats, file uploads and the voice button. It doesn't do any of the thinking.
Chat bubbles, a history sidebar, an upload button, a "new chat" button.
The code running the show behind the window. It writes the hidden instructions, decides which bits of your chat to re-send, and runs the loop that lets the model take several steps in a row.
The bot sticks to company rules, remembers what you said ten minutes ago, and works through a task step by step without you poking it.
Things the harness runs on the model's behalf. The model asks for one, the harness runs it, and the result comes back as plain text.
"Searched the web", "Ran code", or an answer that quotes your own company's documents.
The trained engine. Text goes in and the next text comes out. It lives in a data centre, and you pay for it by how much text it reads and writes.
A model picker dropdown, or a line in your company's AI policy saying which provider you're allowed to use.
It's a new engine dropped into the same plumbing.
Keep the harness and tools exactly as they are and swap out the engine in the middle. Here's Anthropic's Claude line-up, from smallest to biggest. It's the same question every time, so pick one.
The answers are illustrative and written to show the difference, so real outputs will vary. Each model also gets upgraded over time, like Opus 5 becoming Opus 5.5.
Bigger isn't automatically better.
You don't need Opus to tell you what's for lunch, and you probably don't want Haiku running your incident.
Everything else is plumbing.
Memory, live info and actually doing things all come from the harness and the tools, not the model.
I've watched teams blame "the AI" when the real problem was the plumbing around it. Call me biased, but check the harness first.
ChatGPT, Copilot, Claude and your work chatbot all run on this same stack, with different parts inside.