Claude Certification Guide CCAO-F

Curriculum › Product and Model Selection · 12% of the exam

Choosing a model: Haiku, Sonnet, Opus

What you’ll be able to do

  • Explain in plain language how Haiku, Sonnet and Opus differ
  • Choose a model by weighing task complexity, volume, speed and cost
  • Decide when results justify moving to a more or less capable model

What you need to know

Three families, one trade-off

Claude comes in model families that trade capability against speed and cost. The exam tests the trade-off, not version numbers, which change regularly.

  • Haiku is the smallest and fastest family and costs the least. It's built for simple, well-defined tasks at high volume, such as classifying, tagging, short rewrites, extracting a few fields or quick first drafts from a clear template.
  • Sonnet is the balanced middle and the sensible default for most everyday work: drafting reports, summarizing documents, analyzing a spreadsheet, preparing a presentation outline.
  • Opus is the most capable family. It's suited to complex, multi-step reasoning, subtle judgment and long, intricate material, such as weighing competing contract terms, building a multi-scenario plan or untangling a messy problem with many constraints. It is slower and costs more.

Which specific models you can pick, and whether you can pick at all, depends on your plan and your organization's settings. The reasoning stays the same either way.

Match the model to the task, not to your ambitions

The instinct the exam is designed to catch is "use the best model for everything." The best model is the one that meets the quality bar at the lowest cost and wait. Four questions help:

  • How complex is the reasoning? Sorting feedback into five fixed categories is simple. Reconciling three conflicting vendor proposals against a budget and a risk policy is not.
  • How much volume? Hundreds of similar small items favor a fast, low-cost model. Waiting time and cost add up across every item.
  • How much does speed matter? A live task where someone is waiting favors a faster model. An overnight strategy memo can afford a slower one.
  • What does a mistake cost? A wrong tag on an internal survey comment is cheap to fix. A flawed analysis going to the board is not. Higher stakes justify more capability and more human review. A larger model does not replace that review.

Adjust based on results

Model choice is not a one-time decision. A practical pattern is to start with the balanced default, or with the smallest model for clearly simple bulk work, and then check the output:

  • If the smaller model keeps missing nuance, mishandles edge cases or loses the thread on a genuinely hard task, move up a tier.
  • If you're using the most capable model for routine reformatting and the results are no better than a smaller model's would be, move down. You're paying in time and cost for capability the task doesn't use.
  • Before blaming the model, check the prompt. If the instructions are vague or the needed context is missing, a bigger model will just give you a more polished version of the wrong thing. Fix the prompt first (Domain 1 and Domain 7).

Mix models across a workflow

Different steps in one piece of work can use different models. An HR coordinator processing hundreds of open-ended survey responses might use a fast model to tag each comment by theme, then use a more capable model once to write the nuanced summary of themes for leadership. Matching each step to its own needs is better than running the whole job on one tier.

All models can produce confident, plausible errors. Choosing a more capable model lowers the chance of some mistakes on hard tasks, but it doesn't remove the need to verify facts, figures and citations before they go anywhere important.

Key concept

Pick the least costly, fastest model that reliably meets the task’s quality bar. Haiku for simple volume, Sonnet as the default, Opus when the reasoning is genuinely hard.

Exam questions describe a task's complexity, volume, urgency and stakes. The right answer follows those signals. "Always use the most capable model" and "always use the cheapest model" are both wrong as blanket rules.

Practice scenario

ScenarioAn operations analyst has two jobs this week. First, label about 800 short supplier emails as "invoice", "delivery update", "complaint" or "other" so they can be routed. Second, write a recommendation on whether to consolidate from five suppliers to two, weighing cost, delivery risk, contract terms and the complaint history. Her manager suggests using the most capable model for both "to be safe."
Work it through, then open this

These two jobs have different profiles. The email labeling is simple, repetitive and high-volume, with fixed categories, so a fast, low-cost model such as Haiku fits. She should spot-check a sample of the labels, and move up a tier only if accuracy turns out to be poor. The consolidation recommendation calls for multi-factor reasoning with real business consequences, so a more capable model such as Opus, or at least Sonnet, is justified. Even so, she needs to verify the figures and contract details herself before the recommendation goes forward. Using the top model for both wastes time and cost on the labeling and adds no safety, because human review is still the safety step.

Hands-on exercise — Compare tiers on two different tasks

Beginner · 15 min

What you’ll learn

  • Seeing where a smaller model is good enough
  • Seeing where extra capability actually shows
  • Separating model problems from prompt problems

  1. If your plan lets you choose a model, give the same simple task (for example “classify these ten customer comments as positive, negative or neutral”) to a faster model and to a more capable one.

    • Why: Simple, well-defined tasks often come out the same across tiers.
    • You should see: Little or no meaningful difference, with the faster model responding more quickly.

  2. Repeat with a hard task: give both models two conflicting policy excerpts and ask for a recommendation that accounts for both.

    • Why: Complex, multi-constraint reasoning is where capability differences tend to appear.
    • You should see: Differences in how well each model handles the tension between the excerpts. Note which one you’d trust as a first draft.

  3. Take the weaker of the two hard-task answers and, without switching models, rewrite your prompt with clearer goals and context.

    • Why: This tests whether the gap was really about the model or about the prompt.
    • You should see: Some or all of the gap closing, which tells you to fix the prompt before reaching for a bigger model.

Exam traps

Using the most capable model for every task because it is the best

On simple, high-volume work, the top model adds cost and waiting time without a meaningful gain in quality.

Choosing the cheapest model for a high-stakes, complex analysis to save money

Matching the model to the task cuts both ways. Genuinely hard reasoning with real consequences justifies a more capable model.

Assuming a bigger model removes the need to check its output

Every model can produce confident errors. Verification and human review stay in place whatever the tier.

Switching models to fix a problem that is really an unclear prompt

If the instructions or context are missing, a bigger model just produces a more polished wrong answer. Fix the prompt first.

Judging a model on one impressive or one bad answer

Check results across a representative sample of the real work before deciding to move up or down a tier.

Memorizing version numbers instead of the trade-off each family represents

Versions change regularly. The exam tests the durable trade-off: fast and low-cost, balanced, or most capable.

Sources

Quick check

An events manager has used one chat for five weeks to plan a conference. Claude has started suggesting a venue the team rejected weeks ago and using a schedule format she told it to drop. What is the most appropriate action?