Lessons0 of 20 done

Module 4: Beyond chat. Lesson 16 of 20, 8 minutes.

Choosing the right model and tool.

How to pick between large, small, reasoning and open-weight models, and between chat apps, built-in features and the API.

Open a chatbot and look at the top of the screen. There is often a small menu with two or three model names in it, and most people never touch it. The default is fine, they figure, and the names do not mean anything anyway.

The names change every few months, so there is no point memorizing them. But the kinds of model behind the names are stable, and knowing the kinds is enough to make a good choice in about two seconds.

Why there is a menu at all

A company that builds language models usually offers several at once, because one model cannot be the best at everything. The biggest model gives the best answers but takes longer and costs more to run. A smaller one answers in a blink for a fraction of the cost, and for most everyday requests you cannot tell the difference. So the menu is not clutter. It is the provider saying: you know your task, you pick.

The same trade-off shows up everywhere models are sold. Software that adds an AI feature has chosen a model for you, usually a fast one. Developers who build on the API (the application programming interface, the plumbing that lets one program send text to a model and get text back) choose per request. You are making the same decision they make, with a menu instead of code.

Large, small and in between

Three kinds cover almost everything you will see.

  • Frontier models are the largest and most capable a provider has. They handle long documents, subtle instructions and messy problems best. They are also the slowest and the most expensive to run, which is why they come with tighter usage limits.
  • Small, fast models are trained the same way but with fewer parameters (the adjustable numbers inside the network). They are quick and cheap, and perfectly good for rewriting an email, summarizing an article, translating a paragraph or answering a plain question. Many products route simple tasks to small models because nobody notices.
  • Reasoning modes let a model think before it answers. Instead of writing the reply straight away, it works through intermediate steps (sometimes shown, sometimes hidden) and then produces the answer. This helps with math, logic, planning, code with tricky edge cases and anything with several steps that depend on each other. It is slower, and on a simple question it adds nothing but waiting.

A rule of thumb: start small, move up when the answer is not good enough, and turn on reasoning when the task has steps.

Open-weight models you can run yourself

Most models live on a provider's servers. You send text, they send text back, and you never see the model itself. Some providers also publish the trained model as a file: the weights, meaning the billions of numbers that training produced. Anyone can download that file and run it on their own computer or company server. These are called open-weight models.

Open-weight is not quite open source. Open source software comes with the full instructions for building it. Open-weight models usually come with the finished numbers but not the training data or the complete recipe, so you can use and adapt the model without being able to recreate it.

Why would you want one? Privacy, first: text that never leaves your machine cannot be read by anyone else. Control, second: the model will not change under you when a provider updates it. Cost, sometimes, if you already own the hardware. The trade-offs are real. Running a capable model locally needs a reasonably powerful computer, the setup takes patience, and the strongest open-weight models usually trail the strongest closed ones.

Four questions for choosing by task

Before you pick, ask:

  1. How good does this need to be? A first draft you will heavily edit can come from a small model. A contract summary you will act on deserves the frontier one, and a second check.
  2. How fast do I need it? If you are going back and forth twenty times, speed matters more than brilliance. If you are asking once and walking away, let it think.
  3. What does it cost? In a chat app the cost is your usage limit: frontier models and reasoning modes use it up faster. In a business the cost is per token, and it adds up across thousands of requests.
  4. Where can this text go? If the material is sensitive, the question is less about which model and more about which plan and which settings, or whether it should be an open-weight model running inside your own walls. The next lesson covers this.

Chat app, built-in feature or API

The model is one thing. The way you reach it is another, and that choice matters just as much.

A chat app is the general-purpose tool: a blank box, a conversation, often with search, file uploads and memory added around the model. It is the most flexible option and the one where your prompting skill matters most.

A built-in feature is AI inside software you already use: the summarize button in your email, the assistant in your spreadsheet, the autocomplete in a code editor. The vendor chose the model and wrote the prompts. You get convenience and context (it already sees your document) and give up control.

The API is for developers. A program sends text and settings to the model and gets text back, paying per use. This is how companies build their own AI products and automate work at scale. You may never use it directly, but every AI feature you meet was built on one.

Consumer plan or business plan

The same model usually comes in a consumer version and a business version, and the model is not what differs. Two things do.

Data handling. Consumer products often use your conversations to improve future models unless you turn that off in settings. Business and enterprise plans usually exclude customer data from training by default and put that promise in the contract. If you are pasting company material, this is the difference that matters.

Limits and controls. Business plans typically allow more messages, more uploads and longer documents, and give an administrator ways to manage who can use what. Free consumer tiers are the most restricted, and often move you to a smaller model once you hit a cap.

If your employer offers a business plan, use it for work. If not, treat the consumer app the way you would treat any website: read the data settings before you paste anything you would not want kept.

Try it

Open any chatbot that has a model picker. Choose the smallest or fastest model and type: Rewrite this in a friendlier tone: Per my last email, the report is still outstanding. Please advise on timing. Then switch to the largest model and send the same prompt. Compare. For most people the small one is good enough, and it came back faster.

Now try a task with steps: I have one-hour meetings at 9, 11 and 2. I need 90 uninterrupted minutes for writing and a 30 minute lunch somewhere between 12 and 1:30. Lay out my day. Run it with reasoning on and off if your app allows it. Notice which version checks the constraints before answering.

Key terms

Frontier model
The largest and most capable model a provider offers, usually also the slowest and most costly to run.
Reasoning mode
A setting where the model works through intermediate steps before giving its answer, which helps on multi-step problems and slows down simple ones.
Open-weight model
A model whose trained numbers are published so anyone can download and run it on their own hardware.
API
The application programming interface, the connection that lets a program send text to a model and receive text back.
Usage limit
The cap a plan puts on how many messages, files or tokens you can use in a given period.

Check yourself

1.You need to rewrite a three-line email so it sounds warmer. Which choice makes the most sense?
2.What does open-weight mean?
3.Your company offers a business plan for the same chatbot you use at home. What is the most important difference?

Start with a small model, move up when the answer is not good enough, turn on reasoning when the task has steps, and let the plan decide where your data goes.