Which Model for Which Job? A Comparison You Can Actually Use

⚠️ Version note. Model names change every few months; the tiers barely change. This lesson teaches the tiers and how to compare them yourself, so it stays true even when the names don't. Last checked: 25 July 2026.

What you'll have at the end

• A clear mental model of the three tiers every AI provider ships, and what each is for
• A decision table you filled in, matching the jobs you actually do to the right tier
• The habit of running a same-prompt comparison so you can feel the difference yourself

Prerequisite

• AI-01 — What a Language Model Actually Does. You need to know a model is predicting text, not looking things up, before "which model" makes sense.
• Access to any chat AI that lets you pick the model (Claude, ChatGPT and most others do).

Why "which model" even matters

People try one model, get a shallow answer, and conclude "AI is dumb." Often they just used the wrong tier for the job — like judging a car by driving it in first gear. Every major provider ships roughly three tiers of the same family, trading capability against speed and cost.

The names in that table will drift. The three-tier shape is stable, because it reflects a real trade-off — more capability costs more compute, and compute costs money and time.

Open your AI of choice and locate where you switch models — usually near the message box or in a menu at the top of the conversation.

[Screenshot: the model selector open in the real interface, showing the available tier names.]

✅ Check: you can see more than one model name and switch between them.

If it moved: the selector is sometimes under the message box, sometimes in the top bar, sometimes behind a settings/gear icon. Look for the family names ("Opus", "Sonnet", "Haiku", or the GPT equivalents).

Ignore the marketing. Models differ along four axes you can reason about:

1. Capability — how well it handles hard reasoning, long code, and multi-step problems. This is the big one, and it's what the "strongest" tier buys you.
2. Speed — how fast the answer comes back. The fast tier can be several times quicker.
3. Cost — if you pay per use (via the API, AI-15) the strongest tier can cost 10–30× the fast one. In a flat-fee chat app you don't feel this directly, but limits still apply.
4. Context window — how much text it can hold at once (a long PDF, a whole codebase). Bigger windows let you paste more without the model losing the start.

✅ Check: you can name, for a task you have in mind, which of these four matters most.

You can't learn the difference from a table. You learn it by watching the same prompt on two tiers.

1. Take one genuinely hard request — a tricky bit of reasoning, a multi-step plan, or a piece of code with an edge case.
2. Send it to the balanced tier. Read the answer.
3. Switch to the strongest tier and send the exact same prompt. Read that answer.
4. Now send something trivial ("rewrite this sentence more formally") to the fast tier and notice how quick it is and that quality didn't suffer.

[Screenshot: two responses to the same hard prompt, side by side or one after the other, from two different tiers.]

✅ Check: you've seen, with your own eyes, one task where the strongest tier was clearly better, and one where the fast tier was perfectly fine and quicker.

If it moved: if your app doesn't let you re-run the same message on a different model, start a new chat, switch the model first, then paste the identical prompt.

Now turn it into something you'll reuse. For the handful of jobs you actually bring to AI, write which tier you'll reach for and why.

✅ Check: your table has at least five real jobs, each mapped to a tier with a one-line reason.

Common mistakes

• Always using the strongest tier "to be safe." It's slower and, on the API, far pricier — and for most tasks the balanced tier is indistinguishable. Match the tier to the job.
• Always using the fast/free tier and blaming "AI" for shallow answers. If a hard task comes back weak, re-ask it on the strongest tier before concluding anything.
• Judging a model on one prompt. Run the comparison; one sample is noise.
• Ignoring the context window. Pasting a huge document into a small-context model means it silently forgets the beginning. Use a big-context model for big inputs.
• Assuming the newest name is always best for you. "Newest" often means "most capable and slowest." Newest ≠ right for a quick task.

Exercise (25 min)

1. Pick three real tasks you'd genuinely use AI for this week — one easy, one medium, one hard.
2. Run each on two different tiers (same prompt both times).
3. For each task, write one sentence: which tier won, and whether the difference was worth the extra wait/cost.
4. Fill in your decision table (Step 4) with at least five jobs.
5. Try the Model Picker tool (linked below) and check its recommendation against your own for two of the tasks — do you agree?

✅ Finish check: you have a comparison written up for three tasks and a decision table of at least five jobs mapped to tiers.

Cheatsheet

THE THREE TIERS (names drift, shape doesn't)
  Strongest  → hard reasoning, long code, research       (slow, pricey)
  Balanced   → most day-to-day work                       (the default)
  Fast       → simple, repetitive, high-volume            (quick, cheap)

FOUR AXES:  capability · speed · cost · context window

RULE OF THUMB
  Default to Balanced.
  If the answer is shallow → re-ask the SAME prompt on Strongest.
  If the task is trivial/repetitive → drop to Fast.
  Big document? → use a big-context model.

NEVER judge a model on one prompt. Run the same-prompt comparison.

Next lesson: AI-06 — Studying With AI: Feynman, Mock Exams, Error Hunting (L2) Related: AI-01 What a Language Model Actually Does (L1) · AI-02 How to Use Claude (L1) · AI-15 Calling AI From Your Own Code (L3) Tool: Model Picker — describe the task, get the tier