What you will be able to do
- Explain why picking a Claude model means trading cost, speed and quality against each other
- Compare Haiku, Sonnet and Opus by speed, how much usage they consume, and the work each suits
- Pick a starting strategy (cost-first or capability-first) based on what the task needs
- Explain why the strongest tier is not automatically right when volume, latency or quota matter
- Use evaluation and effort level to adjust a model choice instead of guessing
Key concept
Cost per completed task — The right way to judge a model choice is the total cost of getting a task done to the standard you need, not the price of each token. A tier that costs less per token can still cost more overall if it needs more attempts or more work, and the reverse can also happen.
1.Cost, speed and quality pull against each other
Every model choice gives something up. Anthropic puts the problem in one line: "Choosing a Claude model means balancing capabilities, speed, and cost." A tier that reasons more deeply is usually slower and uses more of your budget. A tier that is fast and cheap may not handle the hardest problems. No tier is best on all three at once, so the real question is which of the three matters most for this task.
The tiers are not specialists. You will not find a "finance model" or a "science model": every Claude model is trained to excel in areas like coding, agentic tasks and knowledge work. What changes from tier to tier is how hard a problem they can reliably carry, and what that capability costs in price and speed. So you choose a tier by asking how difficult and how demanding the task is, not what subject it covers.
This matters even if you never see a bill. On Claude plans, a heavier model draws down your usage allowance faster. Anthropic's consumption guidance for Enterprise admins says plainly that model choice has a direct and significant impact on spend. Picking a tier is a budgeting decision as much as a quality one.
Checkpoint 1 of 8· Check yourself
A colleague says: "We do legal analysis, so we need the legal-specialist model tier." What is the most accurate correction?
Anthropic does not split model classes by domain. The difference is how much difficulty each tier can reliably handle, weighed against its price and speed.
“how hard a problem they can reliably carry, and what that capability costs in price and speed”Source: claude.com
Checkpoint 2 of 8· Exam question
An operations lead has two hundred short, standard customer replies to prepare today and is worried about her weekly allowance. What should she do first?
Correct answer: B — Separate the replies that genuinely need judgment from the routine bulk of the batch.
- A. Incorrect. Applying the most expensive tier to two hundred routine items is the fastest way to exhaust the allowance she is worried about.
- B. Correct. Once the small number of awkward cases is identified, the routine bulk can go to the fastest and cheapest tier and the allowance is reserved for the cases that need it.
- C. Incorrect. Spreading the work delays the replies without reducing the total allowance the batch consumes.
- D. Incorrect. An estimate of consumption does not change what the batch costs or tell her which parts need care.
2.Haiku, Sonnet and Opus as trade-offs
With those three axes in mind, the tiers form a ladder. Each step up buys reasoning depth and costs speed and usage. The table below combines Anthropic's comparative latency labels with the token-intensity ratings from its consumption guidance for Claude plans.
| Tier | Comparative latency | Token intensity | Best for |
|---|---|---|---|
| Haiku | Fastest | Low | Simple lookups, summaries, fast responses |
| Sonnet | Fast | Moderate | Everyday tasks, writing, analysis, Q&A |
| Opus | Moderate | High | Complex reasoning, research, writing, analysis, coding, and multi-step tasks |
| Fable | Slower | Very High | The hardest, open-ended work |
In everyday terms: Haiku is the lowest-cost, fastest tier, designed for high-frequency workloads where latency and cost matter. Sonnet balances performance, cost and speed across the widest range of general-purpose work. Anthropic's advice for keeping usage under control is to use Haiku, or the most recent Sonnet, for most tasks, because they perform well at lower token cost than Opus. Opus is for reasoning-heavy work. It is slower and uses more, so you spend it where the depth pays off. Above Opus sits Fable, for the hardest open-ended work. The rule of thumb is to move to Fable only if Opus struggles; if Opus already clears the quality bar, its speed and price may make it the better choice.
Checkpoint 3 of 8· Match them up
Match each tier to the situation it fits best
Tap a term, then the definition that fits it.
Each tier up trades speed and usage for reasoning depth. Haiku is built for speed and low cost, Sonnet for balance, Opus for depth, and Fable for cases where Opus falls short.
“Haiku models are designed for high-frequency workloads where latency and cost matter.”Source: claude.com
3.Two ways to choose a starting tier
Knowing the ladder doesn't tell you where to start on it. Anthropic describes two approaches, and each one fits a different kind of task.
Cost-first: start low and move up only when needed. Begin with a fast, low-cost model, test it thoroughly on your task, check whether the results meet your requirements, and upgrade only if necessary for specific capability gaps. This approach suits prototyping, applications with tight latency requirements, cost-sensitive work, and high-volume, straightforward tasks. It also keeps iteration quick and cheap.
Checkpoint 4 of 8· Put it in order
Put the cost-first approach in the order Anthropic gives it
- 1.Begin with a faster, more cost-effective model
- 2.Evaluate whether performance meets your requirements
- 3.Test your use case thoroughly
- 4.Upgrade only if necessary for specific capability gaps
You start cheap, gather evidence, judge it against your requirements, and move up a tier only when that evidence shows a real capability gap.
“Upgrade only if necessary for specific capability gaps.”Source: platform.claude.com
Capability-first: start strong, then economise. For complex tasks where intelligence matters most, you implement with the strongest starting point for your task, then optimize to more efficient models down the line. In practice: start at the top, tune your prompts for that model, check whether it meets your requirements, and then save cost by lowering effort or moving down a tier as the workflow matures. This approach suits complex reasoning, scientific or mathematical work, tasks that need nuanced understanding, and applications where accuracy outweighs cost.
Starting strong has one more benefit. Anthropic notes that starting with a smaller model can make it harder to tell model failures apart from setup failures. If a weak result comes from a vague prompt or missing context, moving up a tier just makes the same mistake more expensive. Fix the setup first.
Checkpoint 5 of 8· Exam question
A marketing team wants to move its weekly performance report from the deepest-reasoning model to a lighter one to save allowance. What should they do first?
Correct answer: D — Produce next week's report on both tiers and compare them against what the report must deliver.
- A. Incorrect. Waiting for complaints makes readers the quality check, and subtle thinning of the analysis rarely produces a complaint.
- B. Incorrect. That answer is a general claim about models rather than evidence from their own report.
- C. Incorrect. Cutting scope changes the deliverable rather than testing whether the lighter tier can produce it.
- D. Correct. A side-by-side run on their own work, judged against an agreed standard, is the only evidence that tells them whether the cheaper tier is adequate here.
Sources1
4.When the strongest tier is the wrong answer
Defaulting to the deepest-reasoning tier is tempting because it feels safe. Anthropic warns that this cuts both ways: the most capable model can be too expensive at scale, and the least expensive model can fall short on quality. Three constraints push the decision down the ladder.
Volume. Higher volumes may be better suited to lower tiers, especially if evaluations show those tasks are completed satisfactorily. A small extra cost per item, multiplied across thousands of items, eats your allowance quickly.
Latency and deadlines. The higher tiers are slower. A large job run on Opus or Fable takes longer than the same job on Sonnet or Haiku. When the work must finish by a fixed time, or a person is waiting for each reply, speed becomes a requirement rather than a nice extra.
Quota. Higher tiers have higher token intensity, so heavy use of them reaches your usage limits sooner. That leaves less room for the work that actually needs depth.
Checkpoint 6 of 8· Check yourself
A support team wants Claude drafting replies inside a live, customer-facing chat that handles a steady stream of conversations all day. Which starting point best fits Anthropic's guidance?
Latency and volume dominate in high-frequency customer-facing work. Anthropic names Sonnet as often the best choice for it. The higher tiers are slower and use more.
“If the model is involved in high-frequency customer facing workloads, then Sonnet is often the best choice.”Source: claude.com
This is where the key concept matters. Per-token price is the wrong measure, because a more capable model may solve a hard task in fewer attempts. But the effect is not guaranteed in either direction. In Anthropic's own measurements, an upgrade that saved money on one benchmark cost more per task on another, because the newer model did more work there. So neither "cheapest per token" nor "strongest available" is a rule. Compare on cost per completed task, measured on your own work.
5.Adjusting the choice with evidence and effort
Once you have a starting tier, you refine the choice with evidence. Anthropic is explicit that having a good evaluation set is the most important step: test with your real prompts and data, then compare tiers on accuracy, response quality and how they handle edge cases. Public benchmarks are only a rough guide, because strong models can solve almost all of the questions on them (called saturation). Your own real tasks tell you more.
Switching tier is not your only lever. Users can choose how much thinking Claude applies to each response, and higher effort levels consume more tokens than lower ones. Anthropic advises that tuning effort is often a better lever than switching models: lower effort favours speed and cost and accepts some loss of capability. Higher effort isn't automatically justified either. As Anthropic puts it, hard work does not automatically need high effort. The guidance for users is to reserve maximum effort for the most demanding tasks and use lower effort for routine ones.
| Your situation | Do this |
|---|---|
| Costs are too high; quality is fine | Sweep effort down on your current model |
| Quality isn't good enough | If you lowered effort, restore it; otherwise try the next tier up at low effort |
| You are choosing or switching models | Compare on cost per completed task, not per token |
Checkpoint 7 of 8· Check yourself
A team's weekly report on Opus at high effort is good enough, but it uses more of their allowance than they would like. What should they try first?
When quality is fine and cost is the problem, Anthropic's first move is to lower effort on the current model, which often helps more than switching models.
“Tuning effort is often a better lever than switching models.”Source: platform.claude.com
Checkpoint 8 of 8· Exam question
An education lead is choosing a model tier to help draft feedback that will contribute to students' marks. What should she do first?
Correct answer: A — Establish how the feedback will be used and what an error in it would cost a student.
- A. Correct. The quality bar comes from the consequence of the output. Once she knows an error affects a mark, the tier decision follows from that bar rather than from cost alone.
- B. Incorrect. Volume is only half the picture, and choosing on throughput ignores what an error here would cost.
- C. Incorrect. It may well be right, but choosing it before the quality bar is stated is a guess rather than a judgment.
- D. Incorrect. Length is not the standard, and choosing on it rewards verbosity rather than accuracy.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.The deepest-reasoning tier is always the safest choice, so use it whenever quality matters.Why is that wrong?
At volume, or against a deadline, the strongest tier can be too expensive and too slow. Lower tiers are often right when evaluations show they do the task well enough.
Covered in When the strongest tier is the wrong answer
2.A vague or weak answer means you need a higher tier.Why is that wrong?
A weak result may come from the setup (prompt, context) rather than the model. Fix the setup before paying for a higher tier.
Covered in Two ways to choose a starting tier
3.The tier with the lowest price per token is always the cheapest way to get the work done.Why is that wrong?
What counts is cost per solved task. A more capable model can need fewer attempts, so per-token price alone can mislead in either direction.
Covered in When the strongest tier is the wrong answer
4.A difficult task always needs the highest effort setting.Why is that wrong?
Higher effort uses more tokens but does not always improve the result. Test lower settings and keep them where quality holds.
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“Choosing a Claude model means balancing capabilities, speed, and cost.”
↩︎ Cost, speed and quality pull against each other“implement with the strongest starting point for your task, then optimize to more efficient models down the line”
↩︎ Two ways to choose a starting tier“Applications with tight latency requirements”
↩︎ Two ways to choose a starting tier“Applications where accuracy outweighs cost considerations”
↩︎ Two ways to choose a starting tier“having a good evaluation set is the most important step in the process.”
↩︎ Adjusting the choice with evidence and effort“Upgrade only if necessary for specific capability gaps.”
↩︎ Checkpoint“Tuning effort is often a better lever than switching models.”
↩︎ Checkpoint - 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.”
↩︎ Cost, speed and quality pull against each other“Sonnet provides a balance of performance, cost, and speed for the widest set of general purpose use cases”
↩︎ Haiku, Sonnet and Opus as trade-offs“If Opus already clears the quality bar, then its speed and price profile may make it the better choice.”
↩︎ Haiku, Sonnet and Opus as trade-offs“Higher volumes of production may be more appropriate for lower classes of models, particularly if evaluations show those tasks are completed satisfactorily.”
↩︎ When the strongest tier is the wrong answer“which can solve almost all of the questions on the test (often referred to as saturation)”
↩︎ Adjusting the choice with evidence and effort“Starting with a smaller model can also make it harder to distinguish between model failures and setup failures.”
↩︎ Exam trap 2“how hard a problem they can reliably carry, and what that capability costs in price and speed”
↩︎ Checkpoint“Haiku models are designed for high-frequency workloads where latency and cost matter.”
↩︎ Checkpoint“If the model is involved in high-frequency customer facing workloads, then Sonnet is often the best choice.”
↩︎ Checkpoint“Cost-per-task is often lower for more intelligent models, especially at lower effort levels, even if the price-per-token is higher.”
↩︎ Prediction - 3.
“Model choice has a direct and significant impact on spend.”
↩︎ Cost, speed and quality pull against each other“Users can choose how much thinking Claude applies to each response”
↩︎ Adjusting the choice with evidence and effort“Encourage users to reserve Max effort for only the most demanding tasks and to use lower effort for routine tasks.”
↩︎ Adjusting the choice with evidence and effort - 4.https://support.claude.com/en/articles/12429409-manage-usage-credits-for-paid-claude-plansOfficial docs
“Use our most efficient Haiku model, or the most recent Sonnet model for most tasks.”
↩︎ Haiku, Sonnet and Opus as trade-offs“These provide excellent performance at lower token costs than Opus models.”
↩︎ Haiku, Sonnet and Opus as trade-offs - 5.https://platform.claude.com/docs/en/about-claude/models/optimizing-for-cost-and-intelligenceOfficial docs
“The most capable model can be too expensive at scale, and the least expensive model can fall short on quality.”
↩︎ When the strongest tier is the wrong answer“That direction is not guaranteed”
↩︎ When the strongest tier is the wrong answer“Compare on cost per completed task, not per token”
↩︎ Key concept“The most capable model can be too expensive at scale, and the least expensive model can fall short on quality.”
↩︎ Exam trap 1“Compare on cost per solved task, not per token”
↩︎ Exam trap 3“Hard work does not automatically need high effort.”
↩︎ Exam trap 4 - 6.
“optimizing for speed and cost while accepting some reduction in capability”
↩︎ Adjusting the choice with evidence and effort