What you will be able to do
- Describe what Haiku, Sonnet and Opus each trade off in speed, cost and reasoning depth
- Explain why Sonnet, not Opus, is the sensible default for everyday work
- Recognise when a task is hard enough for the gap between tiers to show
- Avoid treating the tiers as subject specialists, or a tier name as a fixed level of ability
Key concept
The difficulty-versus-cost axis — Haiku, Sonnet and Opus share the same broad skills. What separates them is how hard a problem each can reliably handle, and what that extra capability costs you in speed and usage.
1.One family, one axis of difference
New users often assume the three Claude tiers are different products, with one for writing, one for code and one for analysis. They are not. Anthropic's model documentation gives every current model the same baseline: text and image input, text output, multilingual work, vision and tool use. Moving to a different tier does not turn a kind of skill on or off.
The tiers are not tuned to particular subjects either. Anthropic's explanation of its model family says it does not recommend one class for finance and another for science. Every class is trained for coding, agentic work and knowledge work alike.
What actually varies is a single axis. At one end is how difficult a problem a model can reliably see through. At the other is what that capability costs you in speed and usage. Haiku, Sonnet and Opus are three points along that axis. Every difference covered below is a consequence of where a tier sits on it.
2.The three tiers as you experience them
Anthropic's help centre describes the tiers in the terms you actually notice when using them. Opus offers deeper reasoning for harder problems, such as large cross-cutting changes, difficult debugging or architectural decisions, and it uses meaningfully more of your quota. Sonnet is described as fast, capable and cost-efficient. Haiku is the fastest and cheapest option, well suited to quick lookups, simple edits and high-volume runs. The Claude release notes compress the same picture into one line: Haiku for speed, Sonnet for complex tasks, Opus for maximum reasoning power.
| Tier | How Anthropic describes it | Comparative latency | Relative price of the three |
|---|---|---|---|
| Haiku | The fastest model with near-frontier intelligence | Fastest | Lowest |
| Sonnet | The best combination of speed and intelligence | Fast | Middle |
| Opus | For long-running agentic coding and knowledge work | Moderate | Highest |
Read the table in both directions. Moving up from Haiku towards Opus buys reasoning depth and costs speed and usage. Moving down buys responsiveness and headroom, and gives up some depth on the hardest tasks. No tier wins every column, which is why all three exist.
Haiku is not a stripped-down model for trivial chores. Anthropic's model-selection documentation lists it for real-time applications, high-volume intelligent processing, and cost-sensitive deployments that still need strong reasoning. Haiku gets its speed and low price by sitting at a different point on the axis, not by missing a category of capability.
A voice assistant application requires the lowest possible response latency among current Claude models so that spoken exchanges feel natural and immediate. Which model tier has the fastest comparative latency?
Correct answer: A — Claude Haiku, since it is described as having the fastest comparative latency of the current model lineup
- A. Correct. In Anthropic's model comparison, Haiku is listed with the fastest comparative latency among the current-generation models.
- B. Sonnet is described as fast, but the comparison table ranks Haiku's comparative latency ahead of Sonnet's.
- C. Greater intelligence does not translate into lower latency; Opus is listed with moderate comparative latency, slower than both Sonnet and Haiku.
- D. The models differ noticeably in comparative latency, not just in output quality, which is precisely why latency is one of the criteria for choosing among them.
3.Why the middle tier is the default
If the tiers sit on one axis, why is the default the middle tier rather than the top one? Because most everyday work is nowhere near the hard end. Anthropic calls Sonnet its versatile class for everyday tasks. It balances performance, cost and speed across the widest range of general-purpose uses, and the model overview calls it the best combination of speed and intelligence. For a typical draft, summary, rewrite or explanation, Sonnet already does the job well, and it does it faster and more cheaply than Opus would.
So the default is a starting point, not a ceiling. Opus earns its place when a task needs more reasoning than Sonnet reliably provides. Haiku earns its place when speed or volume matters more than depth.
4.Where the gap between tiers actually shows
The task typically takes a lot of time, it involves multiple steps, or it has not been solved before. Any of these puts the task near the hard end of the axis, where tiers stop producing similar results.
Every tier handles the shared baseline, so on tasks well within all three tiers' abilities their answers look alike. The differences appear as a task approaches the hard end of the axis. Anthropic's model-selection documentation describes the same kind of work as suited to starting with the strongest model: complex reasoning, tasks that need nuanced understanding, and applications where accuracy outweighs cost.
The flip side matters just as much. On a simple task a stronger tier adds little, but it still costs more. The help centre says plainly that Opus uses meaningfully more of your quota and suggests switching to Sonnet for routine work. If several people share an allowance or it is limited, routine jobs run on Opus use up budget that the genuinely hard problems then cannot get, and they gain almost nothing in return.
A startup has a new application idea and wants to minimize development cost and iteration time before deciding whether a more capable model is truly needed. Following Anthropic's recommended approach for this situation, which model should they implement first?
Correct answer: A — Claude Haiku, then upgrade only if testing reveals a specific capability gap
- A. Correct. Anthropic recommends starting with a fast, cost-effective model like Haiku, testing thoroughly, and upgrading only if performance reveals a specific capability gap, which fits a startup minimizing cost and iteration time.
- B. Starting with Opus and optimizing down over time is the recommended path for complex reasoning tasks where accuracy outweighs cost, not for a startup explicitly prioritizing low initial cost and fast iteration.
- C. Anthropic frames two valid starting approaches based on the use case; Sonnet being the universal first choice regardless of budget priorities is not part of that guidance.
- D. Model selection should follow the technical criteria of capability, speed, and cost fit for the application, not external non-technical instructions unrelated to those criteria.
5.Tier names describe a role, not a fixed level
No. A tier name marks a position within a generation, not a permanent level of ability. When Haiku 4.5 launched, the release notes called it Anthropic's fastest, most cost-efficient model, and said the new small model matched Sonnet 4 on coding, computer use and agent tasks. The smallest tier of a newer generation can reach what the middle tier of an older one did.
That is why the help centre warns that exact model names, versions and availability change over time. What lasts is each tier's role: fastest and cheapest, balanced default, or deepest reasoning. Learn the roles, not a list of version numbers.
Anthropic's own product teams also choose by role. Claude in Chrome was changed to default to Haiku 4.5 so it would feel faster and more responsive, and users can switch back to Sonnet. In the Claude app, the release notes tell you to switch between Haiku, Sonnet and Opus anytime based on what you need. The tier is a choice you make for each task, not a commitment you make once.
A team is designing an agent that must autonomously produce very long outputs, such as an extensive generated report or a large batch of generated code, and wants the model tier best associated with handling the most demanding, large-scale generation tasks. Which tier should they choose?
Correct answer: A — Claude Opus, since it is positioned for the largest-scale, most complex generation and engineering tasks
- A. Correct. Opus is recommended for the largest-scale, most complex engineering and generation work, matching a scenario centered on demanding, large-scale autonomous output generation.
- B. Lower cost per token is a real advantage of Haiku, but for the most demanding, large-scale generation tasks Anthropic points to the most capable tier rather than defaulting to the cheapest one.
- C. Long-form output generation is supported across the current model lineup, not exclusively by Sonnet, so this option misstates the family's capabilities.
- D. Model tiers differ in intelligence, latency, and cost for demanding generation tasks, so the choice of tier is not interchangeable for the most complex, large-scale work described here.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Each tier specialises in a subject, so Opus is the one for science or finance and Haiku is the one for casual chat.Why is that wrong?
The tiers are not subject specialists. All of them are trained for the same broad kinds of work. They differ in how hard a problem they can reliably handle and what that costs in speed and usage.
Covered in One family, one axis of difference
2.Haiku is a cut-down model that lacks features such as vision or tool use, so it only suits trivial tasks.Why is that wrong?
Every current model shares the same capability baseline, including vision and tool use. Haiku is cheaper and faster because of where it sits on the difficulty axis, not because features are missing.
Covered in The three tiers as you experience them
3.Opus is the strongest tier, so running everything on it is the safe choice.Why is that wrong?
Opus uses meaningfully more quota. On routine work it adds little that Sonnet does not already deliver, so using it by default drains a shared or limited allowance for no real gain.
Covered in Where the gap between tiers actually shows
4.Any Sonnet model is always more capable than any Haiku model.Why is that wrong?
Tier names are relative within a generation. Haiku 4.5 was described as matching Sonnet 4 on coding, computer use and agent tasks.
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.https://platform.claude.com/docs/en/models/overviewOfficial docs
“All current models support text and image input, text output, multilingual capabilities, vision, and tool use.”
↩︎ One family, one axis of difference“The best combination of speed and intelligence”
↩︎ Why the middle tier is the default“All current models support text and image input, text output, multilingual capabilities, vision, and tool use.”
↩︎ Exam trap 2 - 2.https://claude.com/blog/claude-models-explained-choosing-the-best-model-for-your-use-caseSecondary source
“Every Claude model is trained to excel in areas like coding, agentic tasks, and knowledge work.”
↩︎ One family, one axis of difference“how hard a problem they can reliably carry, and what that capability costs in price and speed”
↩︎ One family, one axis of difference“Sonnet provides a balance of performance, cost, and speed for the widest set of general purpose use cases”
↩︎ Why the middle tier is the default“If it typically takes a lot of time, involves multiple steps, or is previously unsolved then a more capable model class is appropriate.”
↩︎ Where the gap between tiers actually shows“how hard a problem they can reliably carry, and what that capability costs in price and speed”
↩︎ Key concept“Every Claude model is trained to excel in areas like coding, agentic tasks, and knowledge work.”
↩︎ Exam trap 1 - 3.
“Opus offers deeper reasoning for harder problems such as large cross-cutting refactors, difficult debugging, or architectural decisions.”
↩︎ The three tiers as you experience them“It is fast, capable, and cost-efficient.”
↩︎ The three tiers as you experience them“Haiku is the fastest and cheapest option, well suited to quick lookups, simple edits, or high-volume scripted runs.”
↩︎ The three tiers as you experience them“It uses meaningfully more of your quota, so consider switching to Sonnet for routine work.”
↩︎ Where the gap between tiers actually shows“Exact model names, versions, and availability change over time.”
↩︎ Tier names describe a role, not a fixed level“It uses meaningfully more of your quota, so consider switching to Sonnet for routine work.”
↩︎ Exam trap 3 - 4.
“Choose between Haiku 4.5 for speed, Sonnet 4.5 for complex tasks, or Opus 4.5 for maximum reasoning power”
↩︎ The three tiers as you experience them“Our latest small model matches Sonnet 4”
↩︎ Tier names describe a role, not a fixed level“Claude in Chrome now defaults to Haiku 4.5”
↩︎ Tier names describe a role, not a fixed level“Our latest small model matches Sonnet 4”
↩︎ Exam trap 4 - 5.
“Real-time applications, high-volume intelligent processing, cost-sensitive deployments needing strong reasoning, sub-agent tasks”
↩︎ The three tiers as you experience them“Applications where accuracy outweighs cost considerations”
↩︎ Where the gap between tiers actually shows