Choosing a Model

The Menu Is Long. Relax.

GPT, Claude, Gemini, Llama, Mistral... the model menu grows monthly, every launch claims to be the smartest, and the leaderboards reshuffle like a deck of cards. Here's the liberating secret: you don't need to pick the "best" model. You need to pick a good-enough model for your task. And that's a much easier question, one that stays answerable even as the specific models change under our feet.

Meet the Families

Think families, not individual models; versions change constantly, but the landscape's shape is more stable:

• GPT (OpenAI): the family that ignited the ChatGPT moment. Broad general-purpose capability, a huge user base and ecosystem of integrations.
• Claude (Anthropic): a family with an emphasis on careful, safety-focused behavior; popular for long documents, writing, and coding. (Full disclosure: a Claude wrote the first draft of this course. It promises it tried to be fair. A human checked.)
• Gemini (Google): Google's family, with strong multimodal chops and natural ties into the Google ecosystem.
• Open-source / open-weight (Llama, Mistral, and friends): models whose weights you can download and run on your own hardware. You handle the setup; in exchange you get control, customization, and privacy that closed APIs can't match.

The Four Questions That Actually Matter

Forget the leaderboard drama. Ask these:

1. What's the task? Hard reasoning, long analysis, and high-stakes writing want a flagship. Autocomplete, tagging, summaries, and chat at volume are happy on small models. Audio and image inputs need a multimodal model. Match the tool to the job, not the hype.

2. What's the budget? API costs are per token, and flagship models can cost many times more than small ones. At one request a day, use whatever you like. At a million requests a day, model choice IS your budget. A common grown-up pattern: route easy queries to a cheap model, escalate hard ones to a flagship.

3. What are the privacy requirements? Using an API means sending data to a provider. For ordinary tasks with a reputable provider's enterprise terms, that's usually fine. For medical records, legal matters, or trade secrets, requirements tighten fast, which is exactly when self-hosted open-weight models shine: the data never leaves your building.

4. How much context do you need? Context windows (Lesson 8!) vary widely across models. Chatting? Almost anything works. Feeding in entire codebases, contracts, or books? The context window may matter more than any intelligence score.

START: What does the task need?
Strict privacy / data cannot leave?  --> open-weight, self-hosted
Analyze huge documents?              --> whichever has the big
context window
Images / audio in the mix?           --> a multimodal model
Hard reasoning, high stakes?         --> a flagship model
High volume, simple task?            --> a small, cheap model
Just chatting / drafting?            --> honestly, almost any
modern model is fine

The Choice in One Table

Actually Deciding: Run a Bake-Off

The leaderboards measure general benchmarks; you care about your task. So audition candidates like you'd hire anyone, with a work sample:

1. Collect 5 to 10 examples of your real task (real support tickets, real documents, real prompts).
2. Run them through two or three candidate models.
3. Compare quality, speed, and price on your examples.
4. Pick the cheapest model that clears your quality bar. That's it. That's the methodology.

And keep perspective on switching costs, which run surprisingly low. Most providers use similar chat-style APIs, so swapping later is usually easy. This isn't a marriage; it's a hire, and re-hiring is cheap.

Model-Picker Quiz