What You Need to Know
This objective names three model families: Haiku, Sonnet and Opus. Each family is a tier: one level of speed, cost and reasoning depth. The tiers make a real trade-off of speed and cost against depth of reasoning. The Haiku family is the lowest-cost tier. Anthropic lists it as the fastest and describes it as built for high-volume, latency-sensitive tasks. A latency-sensitive task is one where the response must arrive quickly. Of the three families that the exam guide names, the Opus family is the most capable reasoning tier, for complex, high-stakes work. The Sonnet family offers the best combination of speed and intelligence: the balanced middle for everyday tasks. No tier is "the good one". Each tier is the right choice for work that matches its strengths.
This lesson uses one made-up example throughout: the head office of a hotel chain, which has two jobs for Claude. One is standardising three hundred hotel descriptions into one website template. The other is reconciling two suppliers' reports that partly disagree into one purchasing recommendation.
The Trade-Off
The most capable tier takes longer for each response. Your choice of model is one of the factors that determine how quickly a usage allowance is consumed. The fastest tier is fast because it is not built for the most demanding work. On Opus, every hotel description would take longer for no gain.
What Capability Means
Capability means depth on demanding work. That depth covers resolving ambiguity and contradiction, multi-step reasoning, synthesis (combining several sources into one conclusion) when the sources partly disagree, and drafting where a subtle error has real consequences. Capability does not mean better at everything, so the top tier is not the universal choice. For routine, well-specified work, the lighter tiers (the faster, lower-cost ones) produce what the task needs, faster. The hotel descriptions are routine. The supplier recommendation is not routine.
Never claim a lighter tier matches a heavier tier's quality on demanding work. The documentation does not claim it. Use a lighter tier because it fits the task and its output is reviewed, never because the quality is equal.
A reviewer can catch errors that are on the page. A reviewer cannot add reasoning that the draft does not contain, except by doing that reasoning herself. Suppose the fastest tier drafts the purchasing recommendation and never compares the two suppliers' delivery terms. The purchasing manager can correct a wrong price that she sees. She cannot correct a comparison that is missing. So review makes a lighter tier safe only when the task is routine.
Reading the Signals in a Task
Lighter-tier signals are high volume, a routine and repetitive shape, simple transformation, speed that matters, and outputs that are individually low-stakes. Examples of simple transformation are reformatting, extracting or standardising. Those outputs are still spot-checked, because the review discipline from Domain 2 never stops. Signals for the most capable tier are ambiguity, contradiction to resolve, judgement with consequences and complex synthesis. Another signal is work where a wrong subtle conclusion costs more than the extra time. When neither set dominates, the balanced middle is the everyday answer.
Size does not decide the tier. The kind of work decides it. Standardising a long list of entries into one template is lighter-tier work however many rows it has. Reconciling two reports that partly disagree into a single recommendation is capable-tier work however short the final page is. The hotel descriptions go to a lighter tier, and a sample of them is checked. The one-page recommendation goes to Opus.
The Model Picker
You choose the model in the chat interface, not in technical configuration. Click the model name and choose another available model. You can change it at any point in a conversation, and the change applies from Claude's next response. The model picker shows the models your account has access to. That list varies by plan and, in workplaces, by what an organisation has enabled.
So the tier decision is not a one-off commitment. Routine preparation can run on a lighter tier, and the hard synthesis can run on the most capable tier, inside one piece of work. The head office can extract each supplier's terms on Haiku, then switch to Opus for the recommendation. Never opening the model picker serves the average task. It serves both the routine end and the demanding end badly. A default model is a setting on your account, not a decision about your work.
In the exam
The guide names the three families. The objective exists because no tier suits every task. It tests whether you read a task's actual demands, not its prestige. That skill remains useful after any change to the set of models. The guide's own sample question describes a high volume of short, routine drafts, where speed and cost matter more than deep reasoning. The right answer there is the faster, lower-cost model. "Always use the top model" is a named wrong answer.
Choosing by a project's importance
"This project is important" does not move a routine task to the top tier. Importance belongs to projects. Demands belong to tasks. The right answer is the tier that matches the current task. A routine reformat inside an important project is still a routine reformat.
Current state (checked 13 November 2026)
Anthropic now offers a fourth model family, Fable, that is more capable than Opus. Its Academy guide to choosing a model says Fable suits "ambitious, long-running, open-ended work". The same guide recommends starting with Opus for most knowledge work. Opus is the default model on the Pro, Max, Ultra and Team plans. The exam guide names only Haiku, Sonnet and Opus, so on the exam Opus is still the most capable tier. Specific model names and the current line-up change over time, and they are on the quick reference.
Exam Traps
Choosing the most capable model for everything because the work is important
Always using the top model wastes money and response time on work that does not need it. Routine high-volume work does not need depth, and a project's importance does not make every task inside it demanding.
Sending work that plainly needs depth to the fastest tier because it is just writing
This saving costs more in the end. Quality demands depend on the cost of being wrong: a short output that carries judgement, contradiction or consequences is exactly where the extra time of the capable tier is worth it, however quick the typing would be.
Accepting a claim that the lighter tier gives the same quality on demanding work
No documentation says a lighter tier matches a heavier one on demanding work; fitness for the task plus review justifies it. The exam guide reserves the most capable model for complex reasoning.
Practice Scenario
An office administrator at a manufacturing company has two jobs side by side. One is standardising six hundred supplier records into one address format. The other is drafting the company reply to a large customer that disputes a contract. Three managers have given partly contradictory advice on that reply. She must keep spending on the routine job as low as she can, without risking the quality of the company position. How should she assign models?
Build Exercise
Match a Week of Tasks to Tiers
What you'll learn
- How to label tasks as routine-at-volume, everyday middle, or high-stakes synthesis
- How to check which models your own account offers in the picker
- How to run a routine task on a lighter tier with a sampled quality check
- How to name what depth gives you on a genuinely demanding task
- List the Claude tasks from your last working week and label each one: routine-at-volume, everyday middle, or high-stakes synthesis
Why: The three labels are the signal sets exam questions use; practising them on your own work makes them fast to spot.
You should see: A short list of real tasks, each with one of the three labels.
- For each task, note the tier its demands call for and the tier you actually used, then open the model picker and confirm which models your account offers
Why: The gap between called-for and actually-used is where the habit of using one model for everything shows itself.
You should see: Two tier columns per task, plus the list of models your picker currently shows.
- Pick one routine task and rerun it on the lighter tier, spot-checking a sample of the output against your quality standard
Why: Fitness for the task plus review is what justifies the lighter tier; the spot-check is what makes the choice professional rather than merely cheap.
You should see: Output that passes your sampled check, produced faster than your usual setting.
- Pick your most demanding recent task and write one sentence on what the capable tier gave, or would have given, that speed could not
Why: Naming the benefit in a sentence tests whether the task genuinely needed depth or just felt important.
You should see: One sentence naming a concrete benefit: a contradiction resolved, a subtle point kept, a consequence avoided.
- Flag any task where you chose by prestige rather than by the task demands, and relabel it with the tier it actually needed
Why: Choosing by prestige is the recurring wrong answer in this objective; catching it in your own week protects you from choosing it in exam questions.
You should see: At least one honest relabel, or a week that genuinely had none.
Sources
- Models overview — Anthropic
- Choosing a Claude model for knowledge workers — Anthropic
- Change the model, effort, and thinking settings — Anthropic
- Claude Certified Associate – Foundations Exam Guide — Anthropic