The tiers, and what moves as you go down
faster and cheaper at the top · more capable and costlier below
- Haikufastest, cheapest, near-frontier quality
- Sonnetthe balance of speed and intelligence
- Opuslong-running, complex, agentic work
How the exam frames it
The task statement names three model types, and that three-way split is the mental model to answer with: Haiku is the speed-and-cost tier, Sonnet is the everyday balance, Opus is the capability tier. Almost every item in this objective is a restatement of that. If a scenario stresses volume, latency or budget, it is pointing at Haiku; if it stresses hard multi-step reasoning, long autonomous work or accuracy on something difficult, it is pointing at Opus; if it describes ordinary professional work, it is pointing at Sonnet.
The documented positioning of the current models, in Anthropic’s own words, is worth reading once. Claude Fable 5.1 is for demanding reasoning and long-horizon agentic work. Claude Opus 5.5 is for long-running agentic coding and knowledge work. Claude Sonnet 5 is described as the best combination of speed and intelligence. Claude Haiku 4.5 is the fastest model, with near-frontier intelligence — which is the phrase to remember, because it contradicts the common assumption that the cheap tier is the weak one.
| Model | Tier | Context window | Price per million tokens |
|---|---|---|---|
| Claude Fable 5.1 | Fable | 1M | $10 in · $50 out |
| Claude Opus 5.5 | Opus | 1M | $4 in · $20 out |
| Claude Sonnet 5 | Sonnet | 1M | $2 in · $10 out |
| Claude Haiku 4.5 | Haiku | 200K | $1 in · $5 out |
Two things in that table repay attention. The spread from the cheapest to the most expensive is an order of magnitude, which is why tier choice is a real budgeting decision for anyone running work at volume rather than a preference. And the context window is not uniform: the larger models document a one-million-token window while Haiku documents two hundred thousand. Token prices matter mostly to people building on the API; in the Claude apps you are spending a usage allowance rather than a per-token bill, but the same ordering holds — heavier models consume it faster.
What actually differs between tiers
It helps to separate four properties that people tend to bundle into the single word “better”.
| Property | What it means | How it moves across the tiers |
|---|---|---|
| Capability | Depth on hard, multi-step or ambiguous problems | Rises from Haiku to Sonnet to Opus and above |
| Speed | How quickly the answer comes back | Haiku is documented as the fastest |
| Cost | Tokens on the API; usage allowance in the apps | Roughly ten times from the cheapest to the dearest |
| Context window | How much can be in front of the model at once | 200K on Haiku; 1M on the larger current models |
Note what is not on that list. Tiers do not differ in what kind of work they will attempt, in tone, in the languages they handle, or in their willingness to follow instructions. All the current models take text and image input, produce text, work across languages and use tools. Choosing Opus does not unlock a feature; it buys depth on problems where depth is what was missing.
What “better” is actually buying
Changes with the tier
- Depth on hard reasoning
- Speed of response
- Cost per unit of work
- Size of the context window
Does not change with the tier
- Whether it knows your company’s figures
- Whether a claim has been verified
- Whether the brief was clear
- Whether it can see a document you did not attach
Choosing in the Claude apps
In the apps the model is chosen from the selector next to the send button — click the model name to switch, and “More models” to see additional options. Alongside it sit two settings that change behaviour within a model, and confusing them with the tier is a common mistake.
- Effort determines how thorough a response is and how much of your usage it consumes. Low and medium suit routine tasks and stretch your usage further; high is the default and is described as the best overall balance of quality and speed; extra high and max exist for long-running and deeply demanding work.
- Thinking is a separate toggle for extended reasoning, showing an expandable section above the answer. On the most capable current models it cannot be turned off at all.
- Administrators on Enterprise plans can restrict which models are available, so the picker is not the same everywhere.
The help centre’s advice on these is refreshingly plain: simple questions, basic information requests and general writing do not need extra effort or thinking, while mathematical problems, coding challenges, project planning and technical analysis are where raising effort or turning thinking on is worth it. In other words, the same discipline as tier choice, one level down.
This objective is about telling the tiers apart. Matching a specific task’s requirements — cost, speed and quality together, with the trade-offs that involves — is objective 3.3, and the context-window consequences of a very long conversation are 3.4.
Traps the wrong answers are built from
| Tempting but wrong | Do this instead |
|---|---|
| Treating the most capable model as the default for everything | Match the tier to what the task needs; volume and latency often point the other way. |
| Assuming the fast tier is the low-quality tier | Haiku is documented as fastest with near-frontier intelligence, not as a weak model. |
| Switching models to fix a vague prompt or a missing file | Check the source and the brief first; a tier change fixes neither. |
| Confusing the effort setting with the model tier | Effort changes thoroughness within a model; the tier changes the model. |
| Assuming every user sees the same model list | Availability varies by plan, and Enterprise administrators can restrict it. |
You should now be able to
- State the defining property of each tier: Haiku speed and cost, Sonnet balance, Opus capability.
- Recognise which tier a scenario is describing from its constraints rather than its adjectives.
- List the properties that change across tiers — capability, speed, cost, context window — and those that do not.
- Distinguish the model tier from the effort and thinking settings inside the apps.
- Explain why a more capable model does not fix a missing source or an unclear brief.
- Account for the gap between the exam’s three-tier framing and the current product lineup.