What you will be able to do
- Compare Haiku, Sonnet and Opus on capability, speed and cost, and say which kind of task each tier suits
- Rank tasks by complexity and stakes, and match each one to a tier and a reasoning depth
- Explain what extended thinking and effort change, and what they cannot guarantee
- Recognise when the faster, lower-cost tier is the right choice
- Use a requirements-first test procedure to pick a tier, mix tiers across the stages of a workflow, and defend the choice against the alternatives
Key concept
Tier-and-effort fit — Choosing a model means trading capability against speed and cost twice. First you pick a tier: Haiku, Sonnet or Opus. Then you pick how hard that tier reasons. The right choice is the cheapest combination that reliably meets the task's quality bar.
1.Quality, speed and cost as one trade
Every Claude tier is a different point on the same trade-off. A more capable tier gives better answers on hard work, but it replies more slowly and costs more for each token. Neither end wins by default. The most capable model can be too expensive at scale, and the cheapest one can fall short on quality. The skill is placing each task at the point that fits it.
| Tier | Comparative latency | Price (input / output) | Built for |
|---|---|---|---|
| Claude Haiku 5.5 | Fastest | From $0.10 / input MTok, from $0.50 / output MTok | High-volume, latency-sensitive tasks such as classification, extraction, and routing |
| Claude Sonnet 5.5 | Fast | $2 / input MTok, $10 / output MTok | The best combination of speed and intelligence |
| Claude Opus 5.5 | Moderate | $4 / input MTok, $20 / output MTok | Long-running agentic coding and knowledge work |
| Claude Fable 5.2 | Moderate | $8 / input MTok, $40 / output MTok | Demanding reasoning and long-horizon agentic work |
The gaps are large. At list price, Opus costs twenty times Haiku's starting rate for each input token, and Sonnet costs half of Opus. Speed moves the other way, from fastest at Haiku down to moderate at Opus. Fable sits above Opus for the hardest long-horizon work. Most everyday decisions, though, come down to Haiku, Sonnet or Opus.
Each tier also has typical work. Haiku covers real-time applications, high-volume processing and subagent tasks. Sonnet covers everyday code generation, data analysis and content creation. Opus covers complex, multi-step work such as large-scale refactoring and complex systems engineering. These are starting points, not rules. Whether a tier fits a particular task is something you confirm by testing, which a later section covers.
Three terms in that list carry much of the tier choice. Agentic work is Claude acting autonomously over many steps, calling tools along the way, instead of answering a single prompt. Long-horizon work is agentic work that runs for hours, such as an agent session or an autonomous coding agent that keeps going until a large job is finished. A subagent is a lower-cost model that a larger session or an orchestrator hands a bulk part of the work to. So Opus's long-running agentic coding means a multihour autonomous job, while Sonnet's everyday work is a single request such as generating a function or analysing a dataset. Subagent tasks suit the cheapest tier because each one is a narrow piece of a bigger job.
Checkpoint 1 of 9· Match them up
Match each Claude tier to the workload it is the natural starting point for.
Tap a term, then the definition that fits it.
Each tier sits at a different point on the quality, speed and cost trade-off. High-volume, latency-sensitive work such as classification goes to the fastest and cheapest tier, Haiku. Volume alone does not make a task complex. Sonnet covers everyday work, Opus covers complex multi-step engineering, and Fable sits above Opus for the most demanding long-horizon work.
“The lowest latency and price | Claude Haiku 5.5 | Real-time applications, high-volume intelligent processing, cost-sensitive deployments needing strong reasoning, subagent tasks”Source: platform.claude.com
2.Matching tier to task complexity and stakes
Complexity is about how much reasoning one item needs, not how many items there are. Sorting 150 expense lines into five named categories is a large job. Each line, though, is a lookup against a list you supplied. That is classification, the work Haiku is designed for, and with a big batch the low price per item matters more, not less.
Further up the ladder, the reasoning gets more open-ended:
- Routine drafting from given material. A two-paragraph status update from five bullet points needs fluent writing but little judgement. This is Sonnet's everyday content-creation work, and Haiku may well do it too.
- Structured comparison with tension. Comparing two vendor proposals against a criteria list and flagging where the trade-offs conflict means weighing several factors that pull against each other. This is a step above drafting.
- Open-ended, high-stakes judgement. Weighing conflicting market evidence and recommending whether to cancel a flagship product line, with the reasoning shown, is complex reasoning that needs nuanced understanding. This is where Opus with deeper thinking earns its cost.
Stakes raise the bar on top of complexity. When an error would reach a board, a client or a regulator, accuracy outweighs cost, and a capability-first starting point is reasonable. A low-stakes internal draft with the same structure can tolerate a cheaper tier, because a weak result costs little to catch and fix.
Checkpoint 2 of 9· Check yourself
Two teams run the same task: compare two vendor proposals against a criteria list and flag conflicting trade-offs. Team A's output goes into a filing to a regulator. Team B's is a rough internal draft for their own planning. How should their starting tiers differ?
Stakes raise the quality bar on top of complexity. If an error would reach a regulator, accuracy outweighs cost and a capability-first start makes sense. A low-stakes internal draft with the same structure can use a cheaper tier, because a weak result costs little to catch and fix.
“Applications where accuracy outweighs cost considerations”Source: platform.claude.com
Checkpoint 3 of 9· Exam question
A manager is matching four tasks to Claude's model tiers, from the fastest, lowest-cost tier up to Claude Opus with extended thinking. Judging by the complexity and reasoning each task demands, which sequence runs from the task best suited to the fastest, lowest-cost tier to the task that most needs Opus with extended thinking?
Correct answer: A — Sort 150 expense lines into categories → Draft client update from bullet points → Compare two vendor proposals against criteria → Recommend on cancelling a product line
- A. The order follows rising complexity and reasoning demand: category sorting is simplest and highest-volume, the bullet-point update needs little reasoning, the proposal comparison weighs conflicting trade-offs, and the cancellation recommendation needs the deepest reasoning from conflicting evidence, so it needs Opus with extended thinking most.
- B. Placing the client update ahead of the expense sort misranks the two simplest tasks. Sorting lines into given categories is the simplest, highest-volume task, so it belongs at the fastest, lowest-cost tier before the drafting task.
- C. Placing the vendor comparison ahead of the client update puts a moderate-to-hard task below a task that needs little reasoning. Weighing conflicting trade-offs against criteria demands more reasoning than turning bullet points into a status update, so the update should rank lower.
- D. Placing the expense sort after the update and the vendor comparison treats a high-volume task as demanding. The sort only applies a provided list of category names, so it is the simplest task and should come first, ahead of drafting and comparing.
3.Extended thinking and effort: reasoning depth within a tier
Choosing a tier is one decision; choosing how hard that tier reasons is another. Extended thinking lets Claude work through a problem step by step before it gives its final answer, which helps on complex tasks. Effort is the setting that controls how much reasoning happens. It runs from low through medium and high up to xhigh and max. Effort affects every token in the response, including text, tool calls and thinking.
| Effort level | What it does | Typical use |
|---|---|---|
| low | Most efficient; significant token savings with some capability reduction | Simpler tasks that need the best speed and lowest costs, such as subagents |
| medium | Balanced approach with moderate token savings | Agentic tasks that need a balance of speed, cost, and performance |
| high | Spends as many tokens as the task needs for excellent results | Complex reasoning, difficult coding problems, agentic tasks |
| xhigh | Extended capability for long-horizon work | Long-running agentic and coding tasks |
| max | Maximum capability with no constraints on token spending | Tasks requiring the deepest possible reasoning |
Changing effort is often a better lever than changing tier. If a tier's answers are good but too slow or too costly, lower the effort on the same tier before dropping to a cheaper one. If the answers are shallow on a hard problem, raising effort is the direct fix; there is no need to prompt around the weakness. Lowering effort has a price, though. The model trades away some capability for speed and cost, and in long multi-step tasks at low effort it is more likely to skip a search, stop early or skip a check.
More thinking does not mean a guaranteed correct answer. Extended thinking gives the model more room to reason, and nothing more. On top models, most of the capability gain is already there at high effort. Above high, the gains are uneven and depend on the task. Max effort runs slower and costs more, often for only a small gain. On some structured-output or less demanding tasks it can even cause overthinking. Opus with extended thinking still produces answers that need checking at the review depth the stakes call for, as covered in review depth and escalation.
Checkpoint 4 of 9· Check yourself
A support team runs a Sonnet-based drafting assistant. Reviewers say the answers are good, but replies are slow and the monthly bill is too high. What should the team try first?
If a tier's answers are good but too slow or too costly, lowering effort on the same tier is often a better lever than changing tier. Lowering effort gives up some capability, so the team should check that quality holds. Effort affects every output token, not only thinking, and max effort costs more, often for a small gain.
“Tuning effort is often a better lever than switching models.”Source: platform.claude.com
4.When the faster, lower-cost tier is the right call
For many tasks the cheaper tier is simply the correct answer. Starting with the faster, lower-cost model is recommended for prototyping, for applications with tight latency requirements, for cost-sensitive work, and for high-volume, straightforward tasks. The principle is to upgrade only when a specific capability gap shows up. Paying for capability the task never uses buys slower replies and a bigger bill, and nothing else.
Equal measured quality settles the question. Suppose blind reviewers rate Sonnet's and Opus's weekly vendor write-ups as equally good against the team's quality bar, and Opus takes noticeably longer. The evidence then favours Sonnet. Opus's extra cost and wait buy nothing that the reviewers could see. The same reasoning applies when a lower tier already meets the bar: its speed and price make it the better choice.
Speed only counts when it is the real constraint. If a manager has two hours to deliver a board pack built on conflicting survey data, an Opus reply that takes a minute does not threaten the deadline. The analysis is the part that will be read closely, so switching to a faster tier would trade away quality to solve a problem the manager does not have. Real latency limits look different: a real-time application, or a live chat where every second of waiting shows.
Price per token can also mislead in the other direction. A more capable model often needs fewer turns and less thinking to get a task right, so its cost per completed task can be lower even though each token costs more. Compare tiers on the cost of a finished, acceptable result.
Checkpoint 5 of 9· Check yourself
An analyst is choosing between Haiku and Opus for a multi-step research summary. On Haiku, each summary takes several extra turns and retries before it is acceptable. On Opus, it is usually right on the first pass. Which comparison should settle the choice?
Price per token can mislead. A more capable tier often needs fewer turns and less thinking to get a task right, so its cost per completed, acceptable result can be lower even though each token costs more. Volume alone does not make a task complex, so batch size does not settle the choice either.
“You are choosing or switching models | Compare on cost per completed task, not per token”Source: platform.claude.com
5.Selecting a tier step by step, and defending it
A requirements-first selection starts with what the task needs, tries the cheapest tier that could plausibly meet it, and moves up only where the evidence shows a gap. For a new recurring task:
- Define the quality bar, and how costly an error would be if it reached the reader.
- Note the speed and volume expectations: how many items run each week, and how fast replies must come.
- Run representative sample tasks, including a few hard ones, on the lowest-cost tier that could plausibly work.
- Score the outputs against the quality bar, and separate capability failures from failures caused by a weak prompt.
- Move up one tier only for the task types that failed, retest them, and record the final tier and when to revisit it.
The order matters. Without a quality bar set first, scoring in step 4 has nothing to measure against. Testing with real prompts and data, rather than imagined examples, is what makes the result reliable; a good set of test cases is the most important part of the process. Step 4's split is just as important. A failure caused by a vague prompt is fixed with prompt diagnosis, not a pricier model. When quality falls short and you have not lowered effort, the next move is the tier above at its default effort. Keep the tests running, because models and prices change.
Some workloads justify starting from the other end: complex reasoning where accuracy outweighs cost can begin capability-first on Opus and optimise downward later. Either way, the defence of a choice has two halves. To say why not a cheaper tier, name the capability failure the cheaper tier showed. To say why not a pricier one, show that the higher tier measured no better at more time and cost. A choice backed only by "Opus is the best model" or "Haiku is cheapest" has neither half.
Checkpoint 6 of 9· Put it in order
A team is choosing a tier for a new recurring task, starting from the cheapest tier. Arrange the selection steps in the sequence that makes the final choice defensible.
- 1.Score the outputs against the quality bar, separating capability failures from failures caused by a weak prompt
- 2.Define the quality bar and how costly an error would be if it reached the reader
- 3.Move up one tier only for the task types that failed, retest them, and record the final tier and when to revisit it
- 4.Run representative sample tasks, including a few hard ones, on the lowest-cost tier that could plausibly work
Testing comes before any upgrade, and the upgrade is limited to the specific gaps the tests reveal. The quality bar comes first because scoring has nothing to measure against without it, and prompt-caused failures are separated out so they get a prompt fix rather than a pricier model.
“Test your use case thoroughly. Evaluate if performance meets your requirements. Upgrade only if necessary for specific capability gaps.”Source: platform.claude.com
Checkpoint 7 of 9· Exam question
A team is choosing a Claude tier for a new recurring task. They want to start from the requirements and try the lowest-cost tier that could plausibly work before moving up. Which sequence should they follow?
Correct answer: D — Define quality bar and cost of errors → Note speed and volume expectations → Run samples on lowest plausible tier → Score outputs and find capability failures → Escalate failing task types and record tier
- A. Speed and volume are noted after the samples have run, so the test did not reflect the weekly item count or reply-time needs. A tier that passes on a few relaxed samples could still be wrong at real volume and speed.
- B. The samples run before the quality bar is defined. Without a fixed bar, the team may set the standard after seeing what the lowest tier produced, which biases the result toward whatever it happened to deliver.
- C. Speed and volume are considered only after scoring. The samples were run and judged without the real operating conditions, so the failures flagged as capability problems may not reflect what the task needs, and the escalation decision rests on incomplete scoring.
- D. The quality bar and the speed and volume expectations are set before any testing, so the samples run under real operating conditions and are judged against a fixed standard. Scoring then comes before any move up, so an escalation is backed by evidence.
6.Assigning different tiers to different stages
One workflow does not need one tier. Pairing a lower-cost model with a more capable one, so that most of the work runs at the lower rate, is a standard pattern. Each stage gets the tier its own complexity calls for.
Take a research workflow with three stages. Extracting dates and figures from 30 short notes is extraction, so it goes to Haiku. Synthesising where those notes contradict each other is open-ended reasoning, and it decides the quality of the whole output, so it goes to Opus. Reformatting the findings into a standard layout is routine and goes back to Haiku or Sonnet. The expensive tier runs only where its capability changes the result.
The usual objections each fail. Using different tiers at different stages does not make results inconsistent: matching tiers to stages is the documented pattern, and each stage is judged against its own quality bar. A human review before release does not make a weak synthesis acceptable, because review catches errors and does not create the analysis that was missing. Giving the hardest stage the fastest tier gets the priority backwards. Turnaround matters least on the stage that needs the most reasoning.
Checkpoint 8 of 9· Match them up
A legal-ops workflow handles 200 supplier contracts in three stages. Assign each stage the tier its own complexity calls for.
Tap a term, then the definition that fits it.
Using different tiers in one workflow is a standard pattern. Extraction and reformatting run at the lower rate, and the capable tier is used only on the synthesis stage, where its reasoning changes the result. Each stage is judged against its own quality bar, so mixing tiers does not make the results inconsistent.
“Multimodel strategies pair a lower-cost model with a frontier model so that most tokens are billed at the lower rate.”Source: platform.claude.com
Checkpoint 9 of 9· Exam question
A researcher has three stages: extract dates and figures from 30 short notes, synthesize where the notes contradict each other, and reformat the final findings into a standard layout. The research is reviewed before release. Which assignment of tiers best balances quality and cost?
Correct answer: C — Use Claude Haiku for extraction, Claude Opus for the contradiction synthesis, and Claude Haiku or Sonnet for the layout stage.
- A. Using Opus for extraction and formatting spends top-tier capability on routine steps, and different tiers across stages do not by themselves make results inconsistent.
- B. Review at the end does not repair a shallow synthesis; the contradiction stage still needs strong reasoning, which the fast tier may not give.
- C. Correct: each stage gets the tier its complexity demands, with fast tiers on routine steps and the most capable tier on the hard reasoning step.
- D. This inverts the logic: the synthesis is the hardest stage and needs capability, while speed is the priority only for the routine ones.
Sources
(6)
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/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.”
↩︎ Quality, speed and cost as one trade“Compare on cost per completed task, not per token”
↩︎ When the faster, lower-cost tier is the right call“If you lowered effort, restore it; otherwise try the next tier up at its default effort”
↩︎ Selecting a tier step by step, and defending it“Run the winner in shadow on a traffic slice before cutover, then keep the suite running.”
↩︎ Selecting a tier step by step, and defending it“Compare on cost per completed task, not per token”
↩︎ Exam trap 4“You are choosing or switching models | Compare on cost per completed task, not per token”
↩︎ Checkpoint - 2.https://platform.claude.com/docs/en/models/overviewOfficial docs
“The best combination of speed and intelligence”
↩︎ Quality, speed and cost as one trade“For high-volume, latency-sensitive tasks such as classification, extraction, and routing”
↩︎ Matching tier to task complexity and stakes“For high-volume, latency-sensitive tasks such as classification, extraction, and routing”
↩︎ Assigning different tiers to different stages“For high-volume, latency-sensitive tasks such as classification, extraction, and routing”
↩︎ Exam trap 2 - 3.
“Speed and capability for everyday coding, agent, and enterprise workloads”
↩︎ Quality, speed and cost as one trade“Agent sessions that run for hours”
↩︎ Quality, speed and cost as one trade“an orchestrator that delegates bulk work to lower-cost workers”
↩︎ Quality, speed and cost as one trade“Tasks requiring nuanced understanding”
↩︎ Matching tier to task complexity and stakes“Applications where accuracy outweighs cost considerations”
↩︎ Matching tier to task complexity and stakes“Tuning effort is often a better lever than switching models.”
↩︎ Extended thinking and effort: reasoning depth within a tier“Applications with tight latency requirements”
↩︎ When the faster, lower-cost tier is the right call“Upgrade only if necessary for specific capability gaps.”
↩︎ When the faster, lower-cost tier is the right call“Create benchmark tests specific to your use case. Having a good evaluation set is the most important step in the process.”
↩︎ Selecting a tier step by step, and defending it“For complex tasks where intelligence and advanced capabilities are paramount, you might want to start capability-first”
↩︎ Selecting a tier step by step, and defending it“Multimodel strategies pair a lower-cost model with a frontier model so that most tokens are billed at the lower rate.”
↩︎ Assigning different tiers to different stages“Choosing a Claude model means balancing capabilities, speed, and cost.”
↩︎ Key concept“Multimodel strategies pair a lower-cost model with a frontier model so that most tokens are billed at the lower rate.”
↩︎ Exam trap 3“Upgrade only if necessary for specific capability gaps.”
↩︎ Exam trap 5“The lowest latency and price | Claude Haiku 5.5 | Real-time applications, high-volume intelligent processing, cost-sensitive deployments needing strong reasoning, subagent tasks”
↩︎ Checkpoint“Test your use case thoroughly. Evaluate if performance meets your requirements. Upgrade only if necessary for specific capability gaps.”
↩︎ Checkpoint - 4.
“Extended thinking gives Claude enhanced reasoning capabilities for complex tasks”
↩︎ Extended thinking and effort: reasoning depth within a tier - 5.
“optimizing for speed and cost while accepting some reduction in capability”
↩︎ Extended thinking and effort: reasoning depth within a tier“In long agent prompts, the model is more likely to skip a search, stop early, or skip a check at low.”
↩︎ Extended thinking and effort: reasoning depth within a tier“It runs slower and costs more than at xhigh, often for only a small gain.”
↩︎ Extended thinking and effort: reasoning depth within a tier“on some structured-output or less intelligence-sensitive tasks it can lead to overthinking”
↩︎ Extended thinking and effort: reasoning depth within a tier“raise effort rather than prompting around it”
↩︎ Extended thinking and effort: reasoning depth within a tier“Most of the gain in capability is already present at high, and above it, gains are uneven and depend on the task.”
↩︎ Exam trap 1 - 6.https://claude.com/blog/claude-models-explained-choosing-the-best-model-for-your-use-caseSecondary source
“Cost-per-task is often lower for more intelligent models, especially at lower effort levels, even if the price-per-token is higher.”
↩︎ When the faster, lower-cost tier is the right call