CertSafari
    Anthropic Claude Certified Associate – Foundations (CCAO-F)· Lessons

    Domain 3 · Lesson 13/30

    Matching Claude model tier to cost, speed and quality

    Align model selection with task requirements (cost, speed, quality)

    16 min read
    3% of exam
    8 sources
    Published 19 Oct 2026
    Docs as of 14 Oct 2026

    What you will be able to do

    • Explain how cost, speed and quality trade against each other when choosing a Claude model tier
    • Start from what a task's output is used for and what an error costs before picking a tier
    • Choose between starting small and moving up, and starting strong and moving down
    • Recognise high-volume or simple work where the strongest tier costs more, runs slower and adds nothing
    • Test a tier change on real work before making it, and decide whether to change the tier or the reasoning effort based on what you see

    Key concept

    The cost–speed–quality trade-off — Every Claude tier and effort level gives you a different mix of capability, response speed and price. No setting wins on all three. The right choice is the cheapest and fastest one that still meets the quality the task really needs.

    1.Cost, speed and quality pull against each other

    Picking a Claude model means trading three things against each other: how capable it is, how fast it answers, and what it costs. They move together. Going up a tier buys more capability, but each response costs more and takes longer. Going down a tier is faster and cheaper, but the model can handle less difficult work reliably. The tiers do not specialise by subject. No tier is the finance model or the science model. They differ in how hard a problem they can reliably carry, and what that capability costs in price and speed. The tiers themselves are covered in the Haiku, Sonnet and Opus lesson. The figures below show how far apart the tiers are.

    Current tiers on speed and price (API list prices per million tokens)
    TierComparative latencyInput priceOutput pricePositioned for
    Claude Haiku 5.5FastestFrom $0.10 / MTokFrom $0.50 / MTokHigh-volume, latency-sensitive tasks such as classification, extraction, and routing
    Claude Sonnet 5.5Fast$2 / MTok$10 / MTokThe best combination of speed and intelligence
    Claude Opus 5.5Moderate$4 / MTok$20 / MTokLong-running agentic coding and knowledge work

    In the Claude apps, you pay in allowance rather than dollars, and the same trade-off applies. Opus uses meaningfully more of your quota. Routine work belongs on a lighter tier, so the allowance is still there when a hard problem comes along.

    Choose the balance from the task, not from the tier. Before comparing models, answer four questions. How hard is the work? Something that takes many steps or has not been solved before needs a more capable tier. How quickly does an answer have to arrive? How many times will the task run? And what will the output be used for, including what a mistake would cost? The last question carries the most weight. A brainstorm that someone rewrites anyway can tolerate rough edges. Feedback that affects a student's marks cannot. It may also need a person to check it, whatever the tier (see when human review is required). Where accuracy matters more than cost, the balance moves toward capability. Where outputs are cheap to check and run in the thousands, it moves toward speed and price.

    Checkpoint 1 of 8· Check yourself

    A legal-tech team says: "Contract summaries are legal work, so we should use the Claude tier that specialises in legal text." What is the better way to choose the tier?

    Checkpoint 2 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?

    Sources123

    2.Two ways to choose a starting tier

    There are two reasonable starting points, and both end the same way: you test, then move. The difference is which direction you move in.

    Start small and move up. Build on Haiku and test the use case thoroughly. Check whether the results meet your requirements, and upgrade only if a specific capability is missing. You iterate quickly and spend little while developing, and for many common applications Haiku turns out to be enough. This approach fits prototypes, applications with tight latency requirements, cost-sensitive deployments, and high-volume straightforward tasks.

    Start strong and move down. Build on Opus, tune your prompts for it, and check the results. Then cut cost over time by lowering effort or moving to a cheaper tier as the workflow matures. This approach fits complex reasoning, scientific or mathematical work, tasks that need nuanced understanding, and applications where accuracy matters more than cost.

    Starting strong is often cheaper than the price list suggests. A more capable model usually needs fewer turns and less thinking to get a hard task right, so the cost of a finished result can come out lower. A small model has a second hidden cost: when it fails, you can't easily tell whether the model fell short or your prompt and setup did. Start small when the work is simple, high-volume and easy to check. Start strong when the work is hard, or when a wrong answer is expensive.

    Checkpoint 3 of 8· Put it in order

    You need accuracy on a complex reasoning task, so you take the start-strong route. Put its steps in the order the guidance gives.

    1. 1.Check whether the results meet your requirements
    2. 2.Build the application on Claude Opus 5.5
    3. 3.Over time, lower the effort or move to a cheaper tier as the workflow matures
    4. 4.Tune your prompts for Opus

    Sources12

    3.When the strongest tier is the wrong answer

    For simple or high-volume work, the strongest tier costs more and runs slower without improving the result. A short standard reply, a product blurb summary or a category label is something a lighter tier already gets right. Paying for deeper reasoning on each item adds cost, not quality. When outputs can be checked, the guidance for high-volume work is to try Haiku first. Large production volumes suit lighter tiers, particularly when evaluations show they complete the task satisfactorily. The same reasoning applies one level up. Claude Fable 5.1 sits above Opus: it is Anthropic's most capable model open to all customers, rated slower than Opus, and priced at $10 per million input tokens and $50 per million output tokens, against $4 and $20 for Opus 5.5. If Opus already meets the quality bar, its speed and price beat Fable 5.1, and the step up is only worth it where Opus still falls short at its highest effort.

    Speed is a requirement too, and it often decides the case. Haiku is the fastest tier and Opus is moderate. An overnight batch of several thousand items moved to the deepest tier 'for quality' can miss a fixed cut-off, and it uses far more budget or allowance getting there. The batch didn't need the extra capability, but it did need to finish on time.

    Most real batches are mixed, so sort them before picking a tier. Separate the few items that need judgment from the routine bulk. Send the bulk to a light tier and the hard cases to a stronger one. This is the idea behind multi-model setups: a lower-cost model handles most of the tokens, and a frontier model gets only the decisions that need it. One pattern has an executor escalate hard calls to an advisor. Another has an orchestrator hand bulk work to cheaper workers.

    If Opus is right for the work but too slow, fast mode is a separate option. It runs the same Opus model up to 2.5x faster at a premium price. It buys speed only: there is no change to intelligence or capabilities, so it does nothing for a task where the model's answer itself is wrong.

    Checkpoint 4 of 8· Check yourself

    Every night, a support team labels 8,000 tickets by category before a fixed 6 a.m. cut-off. About 2% are ambiguous and need real judgment. What is the best setup?

    Checkpoint 5 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?

    Sources42156

    4.Test a tier change on your own work before making it

    Decide on a tier change with evidence from your own task. A good evaluation set is the most important part of choosing a model. Use your real prompts and data, compare the tiers on accuracy, response quality and edge cases, and only then weigh performance against cost. For a recurring report, that means producing the next one on both tiers and checking each against what the report has to deliver. Compare cost per completed task, not price per token. A cheaper tier that needs a rerun or heavy editing may not save anything.

    Three shortcuts look like testing but aren't. Switching straight away and waiting for complaints hands quality control to whoever happens to notice. Asking Claude whether a lighter model could do the job produces a prediction, not a measurement. Shortening the deliverable to suit the lighter tier changes the task, so the comparison no longer tells you whether the lighter tier can do the original job. Published benchmarks also have a limit: strong models score near the top, so the differences that matter only show up on your own workload.

    Checkpoint 6 of 8· Check yourself

    Before moving its weekly market report from Opus to Sonnet to save money, an analyst wants solid grounds for the switch. Which approach gives them that?

    Checkpoint 7 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?

    Sources142

    5.Adjusting tier and reasoning effort on evidence

    Tier is not the only setting. Within one model, the effort level trades thoroughness for speed and cost, and changing effort is often a better move than changing models. Effort applies to every output token: thinking, tool calls and the response text. At higher effort, Claude reads more, checks more and keeps working longer on a multi-step task before replying. At lower effort it uses fewer tokens and will ask for more context sooner.

    Effort levels and where each fits
    LevelWhat it doesTypical use
    lowMost efficient; significant token savings with some capability reductionSimpler tasks that need the best speed and lowest costs, such as subagents
    mediumBalanced, moderate token savings; default on Claude Opus 5.5 and Claude Haiku 5.5Agentic tasks that need a balance of speed, cost, and performance
    highSpends as many tokens as the task needs; default on most other models, including Claude Sonnet 5.5Complex reasoning, difficult coding problems, agentic tasks
    xhighExtended capability for long-horizon workLong-running agentic and coding tasks
    maxMaximum capability with no limit on token spendingThe deepest reasoning and most thorough analysis

    Start at the model's default effort and move it up or down based on what your tests show, not on assumptions. A few common situations:

    • Costs are too high but quality is fine: lower the effort on the model you already use.
    • A task finishes correctly but takes longer than necessary: reduce effort.
    • Quality isn't good enough: if you had lowered effort, restore it; otherwise try the next tier up at its default effort.
    • Outputs can be checked automatically: run at low effort and rerun only the failures at high.

    How a task failed tells you which setting to change. If Claude had all the relevant context, clearly worked through the problem and still got it wrong, the model's capability is the limit, so move up a tier. If it went wrong by skipping a file, not running the tests or stopping a refactor halfway, it did too little work, so raise effort. A vague first draft may be neither: it can be a prompt that is missing context or a clear goal, which is covered in diagnosing underperforming prompts. Changing the tier fixes none of that.

    Checkpoint 8 of 8· Match them up

    Match each observation from your tests to the setting change it calls for.

    Tap a term, then the definition that fits it.

    Sources1748

    Sources

    (8)

    Every claim above is drawn from one of these pages, quoted as it was written on the date shown.

    1. 1.
      “Choosing a Claude model means balancing capabilities, speed, and cost.”
      ↩︎ Cost, speed and quality pull against each other
      “Applications where accuracy outweighs cost considerations”
      ↩︎ Cost, speed and quality pull against each other
      “Upgrade only if necessary for specific capability gaps.”
      ↩︎ Two ways to choose a starting tier
      “This approach allows for quick iteration, lower development costs, and is often sufficient for many common applications.”
      ↩︎ Two ways to choose a starting tier
      “Consider increasing efficiency by lowering effort or downgrading models over time with greater workflow optimization.”
      ↩︎ Two ways to choose a starting tier
      “Claude Fable 5.1 (claude-fable-5-1) is Anthropic's most capable model open to all customers.”
      ↩︎ When the strongest tier is the wrong answer
      “Multi-model strategies pair a lower-cost model with a frontier model so that most tokens are billed at the lower rate.”
      ↩︎ When the strongest tier is the wrong answer
      “having a good evaluation set is the most important step in the process”
      ↩︎ Test a tier change on your own work before making it
      “Test with your actual prompts and data.”
      ↩︎ Test a tier change on your own work before making it
      “Tuning effort is often a better lever than switching models.”
      ↩︎ Adjusting tier and reasoning effort on evidence
      “Test with your actual prompts and data.”
      ↩︎ Exam trap 5
    2. 2.
      “how hard a problem they can reliably carry, and what that capability costs in price and speed”
      ↩︎ Cost, speed and quality pull against each other
      “If it typically takes a lot of time, involves multiple steps, or is previously unsolved then a more capable model class is appropriate.”
      ↩︎ Cost, speed and quality pull against each other
      “Starting with a smaller model can also make it harder to distinguish between model failures and setup failures.”
      ↩︎ Two ways to choose a starting tier
      “If Opus already clears the quality bar, then its speed and price profile may make it the better choice.”
      ↩︎ When the strongest tier is the wrong answer
      “which can solve almost all of the questions on the test (often referred to as saturation)”
      ↩︎ Test a tier change on your own work before making it
      “Higher volumes of production may be more appropriate for lower classes of models, particularly if evaluations show those tasks are completed satisfactorily.”
      ↩︎ Exam trap 1
      “The main difference across model classes is in how hard a problem they can reliably carry, and what that capability costs in price and speed.”
      ↩︎ 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. 3.
      “It uses meaningfully more of your quota, so consider switching to Sonnet for routine work.”
      ↩︎ Cost, speed and quality pull against each other
    4. 4.
      “High-volume work with outputs you can check”
      ↩︎ When the strongest tier is the wrong answer
      “Compare on cost per completed task, not per token”
      ↩︎ Test a tier change on your own work before making it
      “Try lower effort on the model you use now”
      ↩︎ Adjusting tier and reasoning effort on evidence
      “The most capable model can be too expensive at scale, and the least expensive model can fall short on quality.”
      ↩︎ Key concept
      “Compare on cost per completed task, not per token”
      ↩︎ Exam trap 2
      “If you lowered effort, restore it; otherwise try the next tier up at its default effort”
      ↩︎ Exam trap 3
    5. 5.
      “Claude Fable 5.1For demanding reasoning and long-horizon agentic work | $10 / MTok | $50 / MTok”
      ↩︎ When the strongest tier is the wrong answer
    6. 6.
      “Fast mode runs the same model with a faster inference configuration. There is no change to intelligence or capabilities.”
      ↩︎ When the strongest tier is the wrong answer
      “Fast mode runs the same model with a faster inference configuration. There is no change to intelligence or capabilities.”
      ↩︎ Exam trap 4
    7. 7.
      “You can trade off between response thoroughness and token efficiency with a single model.”
      ↩︎ Adjusting tier and reasoning effort on evidence
      “Reduce effort if a task completes but takes longer than necessary, or if you want a faster, more interactive working style.”
      ↩︎ Adjusting tier and reasoning effort on evidence
    8. 8.
      “If Claude has all the pertinent context, clearly tried, and still got it wrong, that's a signal to pick a more capable model.”
      ↩︎ Adjusting tier and reasoning effort on evidence
      “by skipping a file, not running the tests, or bailing on a refactor partway through, pick a higher effort level”
      ↩︎ Adjusting tier and reasoning effort on evidence

    Spotted a mistake, or was something unclear? Tell us.