What you will be able to do
- Rank the current Claude tiers by price and latency class, and recognise which specs do not set them apart
- Translate a workload profile into binding latency, cost, and quality constraints
- Pick the tier that satisfies every binding constraint and explain why each rejected tier fails
- Recognise when latency is not binding and asynchronous batch processing changes the cost picture
Key concept
Binding-constraint tier selection — Treat each stated latency, cost, and quality requirement as a pass/fail gate. Choose the cheapest, fastest tier that passes all of them on your own eval. Quality beyond the bar is not a reason to move up a tier.
1.The four current tiers and what separates them
Model tier selection matters because the tiers differ in price and speed, and those differences multiply across every request a workload sends. The current lineup has four tiers. Claude Fable 5.2 is the most capable model open to all customers. Claude Opus 5.5 is built for long-running agentic coding and knowledge work. Claude Sonnet 5.5 is pitched as the best combination of speed and intelligence. Claude Haiku 5.5 targets high-volume, latency-sensitive tasks such as classification, extraction, and routing. If one tier were best on every axis, there would be no decision to make: the most capable model can be too expensive at scale, and the cheapest can fall short on quality.
| Tier | Comparative latency | Input / output price | Default effort | Context window / max output |
|---|---|---|---|---|
| Claude Fable 5.2 | Moderate | $8 / $40 per MTok | high | 1M / 128K |
| Claude Opus 5.5 | Moderate | $4 / $20 per MTok | medium | 1M / 128K |
| Claude Sonnet 5.5 | Fast | $2 / $10 per MTok | high | 1M / 128K |
| Claude Haiku 5.5 | Fastest | From $0.10 / $0.50 per MTok | medium | 1M / 128K |
The Default effort column refers to the effort parameter, a per-request setting that trades intelligence for latency and cost within a single model. Higher levels buy more reasoning at the price of speed and spend; lower levels do the reverse. The levels run up through high and xhigh to max, with xhigh sitting between high and max, and low is the cheapest and fastest. The column shows where each tier starts when a request sets nothing.
Price falls in a strict order: Fable 5.2, then Opus 5.5, then Sonnet 5.5, then Haiku 5.5. Each step down roughly halves the input and output rate until Haiku, which costs far less. Latency follows a similar order but in coarser classes. Fable and Opus are both Moderate, Sonnet is Fast, and Haiku is Fastest. One consequence is that moving from Opus to Fable buys capability at double the price with no change in latency class.
Several specs are the same on every tier, so they never decide a selection. All four have a 1M-token context window, 128K tokens of max output, and a June 2026 knowledge cutoff. All current models also support text and image input, vision, multilingual use, and tool use. If a scenario's only argument for a tier is a large context window or image input, the argument is empty.
Checkpoint 1 of 7· Match them up
Match each current Claude tier to the workload it is positioned for
Tap a term, then the definition that fits it.
The tiers share context window, max output and modalities, so price, latency class and intended workload are what separate them. Fable and Opus are both Moderate latency, so moving up from Opus to Fable doubles the price without making responses faster. Sonnet is Fast, and Haiku is the Fastest and cheapest tier, built for high-volume work.
“Claude Sonnet 5.5The best combination of speed and intelligence | Claude Haiku 5.5For high-volume, latency-sensitive tasks such as classification, extraction, and routing”Source: platform.claude.com
Checkpoint 2 of 7· Exam question
You are the architect for a support platform that must classify roughly 5 million inbound tickets per day into 12 routing categories, returning only a label. Binding constraints: routing must happen in real time, spend per ticket must stay minimal, and accuracy must be at least 95% on the team's labeled set. On that set Haiku 5.5 scores 97% and Sonnet 5.5 scores 98%. Which tier selection satisfies all binding constraints?
Correct answer: C — Haiku 5.5, because it clears the 95% bar on the labeled set while being the fastest and lowest-priced tier for a narrow, high-volume task.
- A. Sonnet 5.5 meets the accuracy bar but costs more per token than Haiku 5.5 and is slower, and unmeasured 'headroom' is not a stated constraint. Paying for it across millions of tickets buys nothing the requirements ask for.
- B. Opus 5.5 also passes quality, but it is priced well above the cheaper tiers and is not in the fastest latency class. Real-time routing at this volume with minimal spend rules it out.
- C. Haiku 5.5 is documented for high-volume, latency-sensitive work such as classification and routing, and it scores 97% against a 95% requirement. It is the cheapest and fastest tier that satisfies every binding constraint.
- D. Fable 5.2 is the most expensive tier and sits in the moderate latency class. Always-on deep reasoning is unnecessary for a label-only task whose quality bar a cheaper tier already clears.
Checkpoint 3 of 7· Exam question
Which ordering of the current Claude models lists them from the highest to the lowest documented base price per input token?
Correct answer: A — Fable 5.2, Opus 5.5, Sonnet 5.5, Haiku 5.5
- A. The models overview documents Fable 5.2 at the top, then Opus 5.5, Sonnet 5.5, and Haiku 5.5 at the lowest base input price. This is the correct descending order.
- B. Opus 5.5 is priced below Fable 5.2, so placing it first is wrong even though Opus is often assumed to be the premium tier.
- C. Sonnet 5.5 is priced below Opus 5.5, so it cannot sit above it in a descending ordering.
- D. This ordering puts Opus 5.5 above Fable 5.2 and Haiku 5.5 above Sonnet 5.5; Haiku 5.5 has the lowest base price, so both swaps are wrong.
2.Reading a workload as binding constraints
A workload profile usually states three kinds of requirement, and each one is a gate rather than a score to maximise. Latency is a class the tier must reach, such as real time or interactive, or something that can wait overnight. Cost is a ceiling, or an instruction to keep spend per item minimal. Quality is a threshold measured on the team's own data, such as at least 95% accuracy on a labeled set. A tier that clears a gate gets no extra credit for clearing it by more.
Quality evidence must come from your own evaluation, not from a tier's reputation. Build benchmark tests specific to your use case, run them with your real prompts and data, and compare models on accuracy, response quality, and handling of edge cases. A labeled-set score is exactly this kind of evidence. Once a cheaper tier scores above the threshold, the question of whether a stronger tier would be "safer" is already answered.
To run the latency gate, read the tier's comparative latency class against the requirement. A real-time requirement points at the Fastest class: Haiku 5.5 is the tier positioned for real-time applications, and Fable and Opus, both Moderate, do not qualify. The classes are comparative rankings, not published response-time thresholds, so a requirement that sits between real time and overnight, such as interactive, has to be settled by timing the candidate tier on your own prompts. A job nobody waits on has no latency gate at all.
Here is a worked case. A support platform must route about 5 million tickets a day into a fixed set of categories, in real time, at minimal cost per ticket, with at least 95% accuracy. On the labeled set, Haiku 5.5 scores 97% and Sonnet 5.5 scores 98%. Haiku passes all three gates. It clears the accuracy bar, it is in the Fastest latency class, and it is the cheapest tier. Narrow, high-volume classification is also the work Haiku is built for. Each alternative fails at least one gate:
| Tier | Gate it fails | Reason |
|---|---|---|
| Claude Sonnet 5.5 | Minimal cost | 20x Haiku's input rate for prompts up to 100K tokens; the extra accuracy point exceeds a bar already met |
| Claude Opus 5.5 | Minimal cost, real time | 40x Haiku's input rate and only a Moderate latency class |
| Claude Fable 5.2 | Minimal cost, real time | Highest price and a Moderate latency class; deeper reasoning is not needed to emit one label |
| Claude Haiku 5.5 | None | 97% clears 95%, Fastest latency, lowest price |
"Headroom against drift" is the argument for the more expensive tier that sounds most reasonable, and it still fails. Drift is a reason to keep re-running the eval, not to pay 20 times more today for accuracy the constraint does not ask for. If wording drifts and Haiku drops below the bar, the eval will show it, and that is the point to move up a tier. For high-volume work whose outputs you can check, the guidance is to try Haiku 5.5 first and compare tiers on cost per completed task.
client = anthropic.Anthropic()
# For time-sensitive applications, use Claude Haiku 5.5 at low effort
message = client.messages.create(
model="claude-haiku-5-5",
max_tokens=1024,
output_config={"effort": "low"},
messages=[
{
"role": "user",
"content": "Summarize this customer feedback in 2 sentences: [feedback text]",
}
],
)
print(next(block.text for block in message.content if block.type == "text"))Checkpoint 4 of 7· Check yourself
An internal tool pulls invoice fields from uploaded forms while a user waits. The requirements are interactive latency, the lowest possible cost per form, and at least 90% field accuracy on the team's labeled set. On that set, Haiku 5.5 scores 93% and Sonnet 5.5 scores 96%. Which tier should the team deploy?
Each requirement is a pass/fail gate. Haiku 5.5 passes all three, so Sonnet's extra accuracy pays for something no constraint asks for, at 20x Haiku's input rate. If drift is a concern, keep re-running the eval and move up a tier only when Haiku falls below the bar. All four tiers have a 1M-token context window, so context size cannot justify choosing Opus.
“Upgrade only if necessary for specific capability gaps.”Source: platform.claude.com
3.When latency is not binding: quality first, then batch
If nobody is waiting on the response, latency stops being a gate. Two things change. A tier's latency class no longer counts in its favour, and asynchronous processing becomes available. Batch processing costs 50% less for work that can wait up to 24 hours, and the discount applies to both input and output tokens on Opus 5.5, Sonnet 5.5, and Haiku 5.5.
Here quality decides the tier and the deadline decides how it is called. Sonnet 5.5 is cheaper, but it fails the only binding constraint, and no prompt or chunking trick has been shown on the eval to recover 8 points. Opus 5.5 is the cheapest tier with evidence that it passes. Fable 5.2 would also pass, but at twice Opus's rate it adds cost without clearing any additional gate, and batching discounts Fable by the same 50%, so it stays twice as expensive. Haiku's Fastest latency class is irrelevant to a job with all night to run, and nothing on the eval shows Haiku reaching a bar that Sonnet misses. The lowest-cost configuration that passes is Opus 5.5 submitted as a batch, because a synchronous call would give up the 50% discount for speed nobody needs.
Finishing early satisfies a deadline that was never at risk. The binding constraint is quality, and Haiku has no eval result showing it meets the bar. Speed only counts when latency is one of the constraints.
Checkpoint 5 of 7· Check yourself
Every night, a pipeline summarises about 50,000 support transcripts into a report that managers read the next morning. On the team's eval, Haiku 5.5 misses the quality bar by 6 points. Sonnet 5.5 and Opus 5.5 both clear it. Which setup should the team use?
No one is waiting on the results, so latency is not a gate. Quality decides the tier, and Sonnet 5.5 is the cheapest tier the eval shows passing. The job can wait, so batching it gets the 50% discount, which a synchronous call gives up. Haiku's speed does not help, because it fails the only binding constraint. Batching Opus still leaves it at twice the price of batched Sonnet, and quality margin beyond the bar is not a requirement.
“batch processing at 50% off for work that can wait up to 24 hours.”Source: platform.claude.com
Checkpoint 6 of 7· Exam question
A legal-tech team runs a nightly job that analyzes 20,000 contracts and must flag nonstandard indemnity clauses. Results are read the next morning, so there is no interactive latency requirement, and quality is the binding constraint. On the team's eval, Opus 5.5 at high effort meets the quality bar, while Sonnet 5.5 misses it by 8 points. Which configuration satisfies the constraints at the lowest cost?
Correct answer: A — Opus 5.5 submitted through the Message Batches API, since asynchronous processing earns the batch discount and the tier already meets the quality bar.
- A. Opus 5.5 passes the eval, and because nobody waits on results the job can use the Message Batches API, which is documented at 50% off base prices. It satisfies quality and minimizes cost.
- B. Fable 5.2 is documented for work where Opus 5.5 at higher effort falls short, and here Opus already meets the bar. Batching halves its price but it remains the costlier choice for no measured gain.
- C. Haiku 5.5's speed is irrelevant when no latency constraint exists, and it is a lower tier than Sonnet 5.5, which already misses the quality bar. It would fail the binding constraint.
- D. Sonnet 5.5 misses the bar on the team's eval, and chunking is an untested assumption that can also split clauses across boundaries. It leaves the binding quality constraint unmet and forgoes the batch discount.
4.Where to start testing, and when to move tiers
Before any eval exists, there are two ways to pick a first tier, and the workload profile tells you which to use. Haiku-first means starting on Claude Haiku 5.5, testing thoroughly, and upgrading only if necessary for specific capability gaps. It suits prototyping, tight latency requirements, cost-sensitive deployments, and high-volume, straightforward tasks. Capability-first means starting on Claude Opus 5.5, optimising prompts for it, and then lowering effort or moving down a tier as the workflow matures. It suits complex reasoning, scientific or mathematical work, nuanced understanding, and cases where accuracy matters more than cost. Opus 5.5 is the general default when nothing points elsewhere.
Fable 5.2 is an escalation, not a default. Move to it when your evals at xhigh or max effort still fall short on demanding reasoning or long-horizon agentic work, such as agent sessions that run for hours or multistep deep research carried through to a finished document. Fable 5.2 also comes with breaking changes to check before switching. They apply if your code changes the system prompt, the tools, or earlier messages in a conversation, moves a conversation from one model to another, uses the computer_20251124 computer use tool, or uses the advisor tool; code that does none of these is not affected by those triggers.
Effort is often a better lever than switching tiers. When costs are too high but quality is fine, try a lower effort level on the current model before moving down a tier. When quality falls short, restore effort if you lowered it, and otherwise try the next tier up at its default effort. Default effort differs by tier: Fable and Sonnet start at high, and Opus and Haiku start at medium. How effort interacts with thinking is covered in reasoning effort and context lifecycle.
A single tier is not the only answer. Multimodel strategies pair a lower-cost model with a frontier model so that most tokens are billed at the lower rate. In one pattern, an executor escalates hard decisions to an advisor. In the other, an orchestrator hands bulk work to cheaper workers, which is where Haiku's subagent role fits. Splitting work between agents is covered in orchestrator and subagent decomposition. For tier selection, the point is that each role in such a system goes through the same gate test on its own.
Checkpoint 7 of 7· Match them up
A team has a deployed workload and new eval results. Match each situation to the next move the lesson recommends.
Tap a term, then the definition that fits it.
Effort trades intelligence for latency and cost within one model, so it is the first lever to try in either direction. Lower it when cost is the problem, and restore it if that hurts quality. Change tiers only when effort cannot close the gap. Fable 5.2 is an escalation for when Opus at its highest effort levels still falls short, not a default.
“Quality isn’t good enough | If you lowered effort, restore it; otherwise try the next tier up at its default effort”Source: platform.claude.com
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.”
↩︎ The four current tiers and what separates them“Try Claude Haiku 5.5 first”
↩︎ Reading a workload as binding constraints“batch processing at 50% off for work that can wait up to 24 hours”
↩︎ When latency is not binding: quality first, then batch“If you lowered effort, restore it; otherwise try the next tier up at its default effort”
↩︎ Where to start testing, and when to move tiers“Compare on cost per completed task, not per token”
↩︎ Key concept“Compare on cost per completed task, not per token”
↩︎ Exam trap 1“batch processing at 50% off for work that can wait up to 24 hours”
↩︎ Exam trap 3“batch processing at 50% off for work that can wait up to 24 hours.”
↩︎ Checkpoint“Quality isn’t good enough | If you lowered effort, restore it; otherwise try the next tier up at its default effort”
↩︎ Checkpoint - 2.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.”
↩︎ The four current tiers and what separates them“Claude Sonnet 5.5The best combination of speed and intelligence | Claude Haiku 5.5For high-volume, latency-sensitive tasks such as classification, extraction, and routing”
↩︎ Checkpoint - 3.
“$0.50 / MTok for prompts over 100,000 tokens”
↩︎ The four current tiers and what separates them“so the same text counts as approximately 30 percent more tokens than on Claude Haiku 4.5”
↩︎ The four current tiers and what separates them“Claude Haiku 5.5 is built for high-volume, latency-sensitive work such as classification, routing, extraction, and subagent tasks.”
↩︎ Reading a workload as binding constraints“It supports adaptive thinking with the effort parameter, a 1M token context window, and up to 128k output tokens.”
↩︎ Exam trap 2“It supports adaptive thinking with the effort parameter, a 1M token context window, and up to 128k output tokens.”
↩︎ Prediction - 4.
“Several Claude models support an effort parameter that trades intelligence for latency and cost within a single model.”
↩︎ The four current tiers and what separates them“Create benchmark tests specific to your use case. Having a good evaluation set is the most important step in the process.”
↩︎ Reading a workload as binding constraints“Real-time applications, high-volume intelligent processing, cost-sensitive deployments needing strong reasoning, subagent tasks”
↩︎ Reading a workload as binding constraints“Upgrade only if necessary for specific capability gaps.”
↩︎ Where to start testing, and when to move tiers“If your evals at xhigh or max effort still fall short on demanding reasoning or long-horizon agentic work, move to Claude Fable 5.2.”
↩︎ Where to start testing, and when to move tiers“If your code changes system, tools, or earlier messages in a conversation, moves a conversation between models”
↩︎ Where to start testing, and when to move tiers“Multimodel strategies pair a lower-cost model with a frontier model so that most tokens are billed at the lower rate.”
↩︎ Where to start testing, and when to move tiers“Tuning effort is often a better lever than switching models.”
↩︎ Exam trap 4“Upgrade only if necessary for specific capability gaps.”
↩︎ Checkpoint - 5.https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-latencyOfficial docs
“For speed-critical applications, Claude Haiku 5.5 offers the fastest response times while maintaining high intelligence.”
↩︎ Reading a workload as binding constraints - 6.
“50% discount on input and output”
↩︎ When latency is not binding: quality first, then batch