15 Which AI model should I use?

16 Which AI model should I use?

The leaderboard moves every quarter. The right way to think about the choice does not. This chapter gives you a decision rule that survives a model update.

16.1 What you will get from this chapter

A simple recipe for picking a model on a given task.
A short note on what paid and thinking tiers buy you.
A grown-up answer to “is one provider better than the others?”.

16.2 The recipe

For analytical work the question “which model” almost always reduces to four sub-questions. In order:

Quality on your task. Run the same real prompt through two candidates. Compare the outputs on something concrete you can judge.
Cost. How much will the task cost when you do it on real data?
Latency. Is it fast enough for your loop — typing, asking, reading, fixing?
Context window. Can it fit the documents, examples, and schema you want to send?

Pick the cheapest model that passes the quality bar. Move up only when the smaller one fails on something you care about.

This is unglamorous. It also means you stop arguing about benchmarks and start picking based on your work.

16.3 Three classes of model, in plain words

Frontier providers (Anthropic, OpenAI, Google) ship models in two tiers and a wrinkle.

Flagship. The strongest the provider sells, slow-ish and expensive. Examples in this edition: Claude Opus 4.7, ChatGPT 5.5, Gemini 3.0. Use for hard reasoning, long-context tasks, the prompts that really benefit from “think a bit more.”
Workhorse. Cheaper, faster, almost as good for most analytical work. Examples: Claude Sonnet 4.6, ChatGPT 5 / 5-mini, Gemini 3.0 Flash. Default to this tier for most coding, classification, and routine document tasks. You will almost always be fine.
Tiny / Haiku. Smallest, cheapest, fast. Examples: Claude Haiku 4.5, ChatGPT Nano. Good for batch classification, tagging, and any task you will run thousands of times. Cost matters more than quality at scale.

A useful rule of thumb: the workhorse tier is 5–15× cheaper than flagship per token, and the tiny tier is 5–10× cheaper again than workhorse. The difference adds up fast in API work.

16.4 “Thinking” mode

Most providers now ship a thinking or extended-thinking mode where the model deliberates internally before producing the visible answer. Costs more (you pay for the internal deliberation in tokens). Helps on hard reasoning tasks, often dramatically. Helps less on routine ones.

Heuristic: turn it on for the prompt that will set the rest of your project’s direction (research design, identification strategy, regression specification). Turn it off for prompts that are essentially “format this nicely.”

16.5 Free vs. paid

The free tiers of every major provider are surprisingly capable in 2026. The paid tiers ($20/month-ish) buy you:

Bigger context — the difference between “I uploaded a paragraph of the codebook” and “I uploaded the codebook.”
Higher daily message limits.
Access to the agent / code features without random throttling.
Often: the actual flagship model rather than a workhorse.

For going through this book once, paid for a single provider is enough. The book does not assume you have all three.

16.6 Privacy briefly, again

We covered this in Chapter 0e. One sentence here: assume non-zero exposure on free tiers; assume “no training on your prompts” on paid consumer tiers and on API tiers; read the policy of the one provider you actually use. Do not paste real personal data into any of them.

16.7 Which is best?

Honest answer: it depends on the task, and the gaps between flagship models are smaller than the gaps between using a model well and using it badly. As of this edition, the working preferences in this book are:

Day-to-day analytical chat: Claude Opus 4.7 or ChatGPT 5.5 — interchangeable for almost every prompt in this book.
Coding agent in the terminal: Claude Code, because the agent and the model are co-designed.
Bulk classification of text: the workhorse tier of either Claude or ChatGPT.
Long-document tasks (300+ pages of PDF): Gemini, sometimes, when the others struggle with the length.

These are working preferences in 2026. They will change. The decision rule will not.

16.8 What changes between editions

Some things in this chapter that are guaranteed to age: the specific model names, the specific prices, the order of the leaderboard. The yearly spring edition refreshes them. If you are reading this on the website, that is the live version; if you are reading a static copy, check the date on Chapter — versions.

16.9 Where AI helps · Where AI bluffs

Helps. Comparing two models on a real task you give it. (Open both in browser tabs; paste the same prompt; read both answers.) Explaining what a benchmark on a leaderboard is actually measuring.

Bluffs. Asserting one model is “best” without specifying for what. Recommending a paid tier when the free one would have been fine for your use. Citing benchmarks from before its training cut-off as if they are current.

16.10 Keep this with you, not the AI

Which provider you commit to for the duration of this book.
The decision to upgrade tier (workhorse → flagship, free → paid) when a real task fails on the smaller one — not before.
The line between “this answer is wrong because the model is wrong” and “this answer is wrong because my prompt was wrong.” Almost always, it is the second.

16.11 Try this

Pick one prompt you ran in Chapter 0e. Run it again on a cheaper model from the same provider — workhorse if you used flagship, tiny if you used workhorse. Compare. Most of the time the cheaper answer is good enough; if so, that is your default for the rest of the book.

16.12 AI and me

How did AI support me here?
How did AI fail me?
How did AI extend me?

16.13 Where to go next

Glossary of LLM terms for reference, or skip to Prompting basics for the practical follow-up.

--- title: "Which AI model should I use?" --- # Which AI model should I use? The leaderboard moves every quarter. The right way to think about the choice does not. This chapter gives you a decision rule that survives a model update. ## What you will get from this chapter - A simple recipe for picking a model on a given task. - A short note on what *paid* and *thinking* tiers buy you. - A grown-up answer to *"is one provider better than the others?"*. ## The recipe For analytical work the question *"which model"* almost always reduces to four sub-questions. In order: 1. **Quality on your task.** Run the same real prompt through two candidates. Compare the outputs on something concrete you can judge. 2. **Cost.** How much will the task cost when you do it on real data? 3. **Latency.** Is it fast enough for your loop — typing, asking, reading, fixing? 4. **Context window.** Can it fit the documents, examples, and schema you want to send? Pick the cheapest model that passes the quality bar. Move up only when the smaller one fails on something you care about. This is unglamorous. It also means you stop arguing about benchmarks and start picking based on your work. ## Three classes of model, in plain words Frontier providers (Anthropic, OpenAI, Google) ship models in two tiers and a wrinkle. - **Flagship.** The strongest the provider sells, slow-ish and expensive. Examples in this edition: Claude **Opus 4.7**, ChatGPT **5.5**, Gemini **3.0**. Use for hard reasoning, long-context tasks, the prompts that really benefit from "think a bit more." - **Workhorse.** Cheaper, faster, almost as good for most analytical work. Examples: Claude **Sonnet 4.6**, ChatGPT **5 / 5-mini**, Gemini **3.0 Flash**. *Default to this tier for most coding, classification, and routine document tasks.* You will almost always be fine. - **Tiny / Haiku.** Smallest, cheapest, fast. Examples: Claude **Haiku 4.5**, ChatGPT **Nano**. Good for batch classification, tagging, and any task you will run thousands of times. Cost matters more than quality at scale. A useful rule of thumb: the workhorse tier is **5–15× cheaper** than flagship per token, and the tiny tier is **5–10× cheaper again** than workhorse. The difference adds up fast in API work. ## "Thinking" mode Most providers now ship a **thinking** or **extended-thinking** mode where the model deliberates internally before producing the visible answer. Costs more (you pay for the internal deliberation in tokens). Helps on hard reasoning tasks, often dramatically. Helps less on routine ones. Heuristic: turn it on for the prompt that will set the rest of your project's direction (research design, identification strategy, regression specification). Turn it off for prompts that are essentially "format this nicely." ## Free vs. paid The free tiers of every major provider are surprisingly capable in 2026. The paid tiers ($20/month-ish) buy you: - Bigger context — the difference between *"I uploaded a paragraph of the codebook"* and *"I uploaded the codebook."* - Higher daily message limits. - Access to the agent / code features without random throttling. - Often: the actual flagship model rather than a workhorse. For going through this book once, paid for a single provider is enough. The book does not assume you have all three. ## Privacy briefly, again We covered this in [Chapter 0e](00e-ai-account.qmd). One sentence here: assume non-zero exposure on free tiers; assume "no training on your prompts" on paid consumer tiers and on API tiers; read the policy of the one provider you actually use. Do not paste real personal data into any of them. ## Which is *best*? Honest answer: it depends on the task, and the gaps between flagship models are smaller than the gaps between *using a model well* and *using it badly*. As of this edition, the working preferences in this book are: - **Day-to-day analytical chat:** Claude Opus 4.7 or ChatGPT 5.5 — interchangeable for almost every prompt in this book. - **Coding agent in the terminal:** Claude Code, because the agent and the model are co-designed. - **Bulk classification of text:** the workhorse tier of either Claude or ChatGPT. - **Long-document tasks** (300+ pages of PDF): Gemini, sometimes, when the others struggle with the length. These are working preferences in 2026. They will change. The decision rule will not. ## What changes between editions Some things in this chapter that are *guaranteed* to age: the specific model names, the specific prices, the order of the leaderboard. The yearly spring edition refreshes them. If you are reading this on the website, that is the live version; if you are reading a static copy, check the date on [Chapter — versions](../versions.qmd). ## Where AI helps · Where AI bluffs **Helps.** Comparing two models on a real task you give it. (Open both in browser tabs; paste the same prompt; read both answers.) Explaining what a benchmark on a leaderboard is actually measuring. **Bluffs.** Asserting one model is "best" without specifying for what. Recommending a paid tier when the free one would have been fine for your use. Citing benchmarks from before its training cut-off as if they are current. ## Keep this with you, not the AI - Which provider you commit to for the duration of this book. - The decision to upgrade tier (workhorse → flagship, free → paid) when a real task fails on the smaller one — *not* before. - The line between *"this answer is wrong because the model is wrong"* and *"this answer is wrong because my prompt was wrong."* Almost always, it is the second. ## Try this Pick one prompt you ran in [Chapter 0e](00e-ai-account.qmd). Run it again on a *cheaper* model from the same provider — workhorse if you used flagship, tiny if you used workhorse. Compare. Most of the time the cheaper answer is good enough; if so, that is your default for the rest of the book. ## AI and me 1. How did AI **support** me here? 2. How did AI **fail** me? 3. How did AI **extend** me? ## Where to go next [Glossary of LLM terms](04-glossary.qmd) for reference, or skip to [Prompting basics](04a-prompting-basics.qmd) for the practical follow-up.