Home › Study guides › CCAO-F › Domain 3 › Lesson 3.2
CCAO-F · Domain 3 · 12% of the exam · Lesson 3.2 · 18 min read
Haiku, Sonnet and Opus: how the Claude models differ
What sets Claude's Haiku, Sonnet and Opus apart, how to read a model's name and version, where you choose one, and how effort and thinking add depth.
Written against objective 3.2 of the official CCAO-F exam guide (Version 1.0, effective July 2026). An independent resource, not affiliated with Anthropic; the practice questions are written from scratch.
3.2.1 Why there is more than one Claude
Owen runs customer-support operations at Settle & Sway, an online furniture retailer. He uses Claude most days and has never touched the model setting: whatever a new chat starts on is what he uses. Then, in a team meeting, a colleague mentions that she sorts reviews "on Haiku" and "saves Opus for the hard stuff". Owen realises he doesn't know what those names mean, or whether his own work would come out differently if he changed them.
Three jobs are on his list this week. The merchandising team wants 50 new product reviews tagged by topic. His twelve agents expect the weekly support summary. And he has to explain why returns of the Linden sofa have tripled since spring, working from 90 return notes that warehouse staff wrote in a hurry. So Owen runs a small experiment: the SAME three requests on each model, noting the speed, quality and depth of every answer.
Before the results, one plain definition. A model is the trained AI system that reads your prompt and writes the reply; every answer in the Claude app comes from one. Anthropic describes Claude as a family of models. The family comes in tiers, and the three the exam names are Haiku, Sonnet and Opus. The exam calls them model types. From your side all three work the same way, in the same chat, with the same kind of prompt. What differs is the trade each tier makes between how much it can reason through and how quickly and cheaply it answers.
One family, three tiers
3.2.2 Three tiers, one trade-off
Here is the belief that trips people up: the most capable model must give the best answer every time, so why use anything else? For a hard problem, capability matters a great deal. For an easy one, it mostly doesn't. Tagging a review as "delivery" or "comfort" has one right answer, and every tier can find it. A more capable model only adds a slower reply and a bigger bite out of your usage allowance, the amount of use most Claude plans include before you have to wait for a reset.
That is the trade the tiers are built around. Haiku is the fastest and lowest-cost tier, made for simple, well-defined work, often in high volume. Sonnet sits in the middle; Anthropic's model overview calls it the best combination of speed and intelligence, and it handles most professional drafting and analysis. Opus is the most capable of the three, for deep reasoning, complex multi-step problems and nuanced synthesis: pulling scattered, messy material together into one sound conclusion. It is also the slowest of the three, and each reply draws more of your usage allowance.
Settle & Sway makes the same kind of choice every day in its warehouse. A replacement cushion cover goes out by parcel courier: fast, cheap and perfectly adequate. A sofa goes on the two-person van, which costs more and has to be booked days ahead. Nobody books the van for a cushion cover, and nobody tries to post a sofa. The van isn't the "better" service; it's the service for heavy loads. Each tier, likewise, is right for a weight of work.
| Tier | Built for | What you trade |
|---|---|---|
| Haiku | Simple, well-defined tasks with clear rules, often in volume: tagging, sorting, short extractions | Less depth on nuanced problems, for the fastest and cheapest replies |
| Sonnet | Most everyday professional work: drafting, summarising, standard analysis | The balanced middle: strong quality at a good speed |
| Opus | Deep reasoning, nuanced synthesis, complex multi-step problems | The slowest of the three and the heaviest on usage, for the most capability |
Memorise the order and what each tier trades. These descriptions stay true as new versions of each model come out.
3.2.3 What changes when you switch tiers
Descriptions only go so far. Owen wants to see the difference on his own work, so he opens a fresh chat for each tier, picks the model and sends the same three requests. The fresh chat matters. In a shared conversation, each tier would see the answers the previous one gave, and the comparison would stop being fair.
The returns request is the one where depth matters most. Look at the two paragraphs just above the pasted notes, which ask Claude to separate what the notes show from what it is guessing, and to name the data that would settle the question.
I run customer-support operations at an online furniture retailer. Our returns report shows that returns of the Linden three-seater sofa have roughly tripled since March. I need to understand why before Friday's operations meeting.
Below are 90 return notes written by warehouse staff. They are short and inconsistent: some give a reason, some only describe the sofa's condition, some both.
Task: group the reasons for return, say which reasons have grown since March, and suggest the most likely explanations for the rise.
Output: a short table of reasons with rough counts, then your explanations in order of likelihood.
For each explanation, say whether the notes show it directly or whether you are inferring it.
Then say what extra information would confirm or rule out each explanation.
<return_notes>
(the 90 notes, pasted in full)
</return_notes>
The same three requests on three tiers
Tag 50 reviews
Haiku is enough
Weekly summary
Sonnet fits
Sofa returns
Opus earns its cost
The first two jobs went the way the tier descriptions predict: the same tags from every tier, and a summary where Sonnet already gave Owen everything he needed. The sofa returns are where the tiers pulled apart. Opus connected three clues the others had left separate: flat seat cushions, a fabric colour that no longer matched the website, and the timing. It suggested that the cushion foam or the fabric batch had changed in March.
Notice what Opus did NOT do. It couldn't know whether a supplier had changed, because nothing in the notes said so; it offered a hypothesis and named the data that would test it. Owen asked the buying team, who confirmed a new foam supplier from March. And the figure "tripled" came from the returns report, not from a model counting notes by eye. A higher tier reasons more deeply about what is in front of it. It doesn't add missing facts, and it doesn't turn a rough count into a checked one.
3.2.4 Names, versions and where you choose
When Owen opens the model selector, the menu in the chat where you pick a model, he finds more than three entries, and each carries a number. The list and the numbers change as Anthropic releases new models. That is less confusing than it looks, because a model's name has three parts.
Reading a model name
The version marks a generation of a tier. When Anthropic releases a new version, it becomes the current model for that tier; older versions stay available for a while and are then retired. The tiers are not updated in step, so their numbers don't line up: at the time of writing, the current Haiku carries a lower version number than the current Sonnet and Opus. Read nothing into that. A version number only compares generations within one tier, never one tier with another, which is why the tiers are worth learning and the numbers are not.
You can change the model at any point in a conversation, and the new one takes over from Claude's next reply. Which models you see depends on your plan and, at work, on your organisation. On the Free plan, for example, the selector offers Haiku and Sonnet, and Opus comes with the paid plans.
On Enterprise plans, administrators can also set the model that new chats start on and turn some models off for particular roles. You can still switch within a chat, among the models you have. If a colleague sees a model you don't, a different plan or role setting is the likely reason.
3.2.5 Effort and thinking: more depth from the same model
Owen's returns puzzle raises one more question: is a bigger model the only way to get a deeper answer? It isn't. Two settings sit beside the model choice and change how hard the chosen model works on each reply. Effort controls how much thinking Claude applies to a response. Thinking lets Claude break the problem down, plan and explore approaches before it answers, and shows a summary of that reasoning in a section you can expand and read.
Think of the difference between asking a different colleague and giving the same colleague more time. Switching tier is the first: another model, with a different balance of capability, speed and cost. Raising effort is the second: the same model, taking longer over each reply. Both buy depth, and both cost something. The Help Center puts it plainly: higher effort gives more thorough responses, but they take longer and use up your usage allowance faster.
| Setting | What you get | Good for |
|---|---|---|
| Low or medium effort | Quicker, lighter replies that stretch your usage further | Routine tasks |
| High effort | What the Help Center calls the "best overall balance of quality and speed" | Work that needs care but not the deepest analysis |
| Max effort | The most thorough reasoning, with the longest waits and the heaviest usage | The hardest analysis, where correctness matters most |
| Thinking | Claude reasons before answering, with a summary you can read | Problems that need planning, and checking how Claude reached a conclusion |
Remember the direction rather than the labels, which vary from model to model. Each model has a recommended effort level, marked as its default, and the Help Center's advice is that the defaults work well for everyday tasks. For complex tasks, such as detailed document analysis, raise the effort, turn on thinking, or both.
Not every model offers both settings. At the time of writing, Haiku has a thinking switch but no effort setting, while the current Sonnet and Opus always think, so on those effort is the dial you turn. Anthropic's guide to choosing a model adds a point worth keeping: adjusting effort is often a better lever than switching models. Owen tries it on the sofa returns. On Sonnet at a higher effort, the reply takes longer and connects more of the clues than it did at the default.
3.2.6 The exam traps
The wrong options on this objective usually push the tier to one extreme, or blame the model for a problem no model can solve.
- ✗ Using Opus for everything "to get the best quality". ✓ Match the tier to the work. On simple, well-defined tasks Opus gives much the same answer as Haiku, only slower and at a higher cost in usage.
- ✗ Using Haiku for everything to save usage. ✓ Keep Haiku for simple, high-volume work with clear rules. Nuanced analysis and synthesis need Sonnet or Opus.
- ✗ Moving to a higher tier when an answer is vague because the prompt lacked facts. ✓ Add the missing material. Claude only knows what is in front of it, whichever tier reads the prompt.
- ✗ Trusting figures because the most capable model produced them. ✓ Work out counts and calculations in a spreadsheet or another structured analysis, and check them. Confidence is not evidence, at any tier.
- ✗ Treating version numbers as the thing to learn. ✓ Learn what each tier trades. Versions change often; the tiers' roles don't.
Four tempting habits, one real choice
3.2.7 Put it together: run your own tier test
You now have the whole picture: three tiers of one family, a gap that shows on work needing judgement, names you can read, and effort and thinking as a second dial. Running Owen's experiment on your own work turns those descriptions into something you have seen.
Knowing the tiers is the first half of model choice. Matching a model to a specific task (3.3) turns it into a decision: weighing cost, speed and quality for the work in front of you, and testing whether a lighter tier is good enough. Context and memory (3.4) cover the other limit that shapes a long piece of work: how much one conversation can hold, and when to restart, summarise or save what matters.
Key takeaways
- ✓ Haiku, Sonnet and Opus are tiers of one Claude family, each trading capability against speed and cost.
- ✓ Haiku fits fast, simple, well-defined work, often in volume; Sonnet fits most everyday professional work; Opus fits deep reasoning and nuanced synthesis.
- ✓ The gap between tiers shows on work that needs judgement; on simple tasks, a higher tier adds time and usage, not quality.
- ✓ No tier supplies missing facts or replaces checking the numbers.
- ✓ A model name is family, tier and version; versions change while the tiers' roles stay the same, and your plan and organisation decide which models you see.
- ✓ Effort and thinking add depth within one model at the cost of time and usage, and the defaults suit everyday tasks.
Check your understanding
4 questions written for this lesson, then one from the CCAO-F question bank on the same topic. Every answer option is explained, including the ones you did not pick. Nothing is stored.
42 CCAO-F questions on Domain 3, free
Every question in the bank is tagged to a domain, so you can drill 42 questions on Product and Model Selection alone, or sit the full 60-question timed simulator.
Open the CCAO-F question bank → Back to Domain 3 →
The question bank is free. It asks for an account only because the quiz engine has to store answers to score them and show which domains are weak. The questions on this page need nothing.