What you will be able to do
- Rule out model tiers using hard limits such as latency and budget before comparing quality
- Decide whether to start cheap-first on Claude Haiku 5.5 or capability-first on Claude Opus 5.5, based on what a failure costs
- Try a different effort level on the current model before paying for a higher tier
- Back a tier choice with your own eval results, compared on cost per completed task
- Plan for behavior changes from new model IDs, infrastructure updates and retirements, using a regression eval suite and a shadow rollout
Key concept
Eval-gated tier selection — Pick the model tier and effort level that meets your measured success criteria on your own task set for the lowest cost per completed task. Run that same suite again whenever the model under your application changes.
1.Comparing tiers against hard limits
Choosing a model tier means trading capability, speed and cost against each other, and every request you send can make that trade differently. The current lineup runs from Claude Haiku 5.5, the fastest and cheapest, through Claude Sonnet 5.5 and Claude Opus 5.5, up to Claude Fable 5.2, the most capable model open to all customers. All four have a 1M-token context window and a 128K-token max output. For current models, the decision therefore rarely depends on context size. It depends on latency, price and how hard the task is.
| Model (API ID) | Comparative latency | Price per MTok (input / output) | Default effort | Built for |
|---|---|---|---|---|
| Claude Haiku 5.5 (claude-haiku-5-5) | Fastest | From $0.10 / from $0.50 | medium | High-volume, latency-sensitive tasks such as classification, extraction, and routing |
| Claude Sonnet 5.5 (claude-sonnet-5-5) | Fast | $2 / $10 | high | The best combination of speed and intelligence |
| Claude Opus 5.5 (claude-opus-5-5) | Moderate | $4 / $20 | medium | Long-running agentic coding and knowledge work |
| Claude Fable 5.2 (claude-fable-5-2) | Moderate | $8 / $40 | high | Demanding reasoning and long-horizon agentic work |
Begin with the limits that cannot move. Suppose a routing call must return in well under a second, or finance has set a fixed budget per task. Each of those limits rules out tiers before you measure any quality. Write them down as numbers next to the quality target. An acceptable response time and a budget per call are success criteria in the same way accuracy is, and most applications need several criteria evaluated together. A target such as "95% of responses under 200 ms" can be tested. "Fast enough" cannot.
Fast mode improves speed without changing tier. It delivers up to 2.5x higher output speed on supported models at premium pricing, which is twice the standard token prices. You pay for the extra speed in money, not in capability. That makes it the right option when a tier is good enough but too slow, and the wrong one when the problem is quality.
Checkpoint 1 of 9· Check yourself
Your eval suite has an accuracy target and a latency target. Claude Sonnet 5.5 meets the accuracy target but misses the latency target. Claude Haiku 5.5 is fast enough but misses the accuracy target. The budget allows a higher price per token. Which change fits best?
Sonnet 5.5 already meets the quality target, so the only thing to fix is speed. Fast mode makes output faster on the same tier and costs more money, not capability. Opus 5.5 has moderate latency, which is slower than Sonnet. Haiku breaks the accuracy target. Raising effort does not reduce latency.
“Fast mode delivers up to 2.5x higher output speed at premium pricing on supported models.”Source: platform.claude.com
2.Two ways to pick a starting tier
There are two sound starting points. The right one depends on what a wrong answer costs you.
Cheap-first starts on Claude Haiku 5.5. You test the use case thoroughly, check whether it meets your requirements, and upgrade only if you find specific capability gaps. This approach suits prototyping, tight latency requirements, cost-sensitive deployments and high-volume, straightforward tasks. Sorting incoming tickets into a fixed set of queues is a good example: the volume is high, there is a hard latency ceiling, and the answer is one of a closed set of labels you can score. With a labelled sample available, run the smallest tier against it and let the accuracy number decide whether to move up.
Capability-first starts on Claude Opus 5.5. You tune prompts for that model and evaluate. Over time, you cut cost by lowering effort or moving down a tier wherever the evals show quality holds. Effort is a model parameter that trades intelligence for latency and cost within a single model; its levels run from low through medium and high to xhigh and max. If evals at xhigh or max effort, the two highest levels, still fall short on demanding reasoning or long-horizon work, move up to Claude Fable 5.2. This approach suits work where accuracy matters more than cost: complex reasoning, scientific work, advanced coding and agents that run with a lot of autonomy. Take an agent that refactors code for hours and opens pull requests that engineers must redo when they are wrong. Each bad run costs engineer time as well as tokens, so you start where quality is highest and work down only where the evidence allows it.
Multimodel designs are the exception to picking a single tier. They pair a lower-cost model with a frontier model so that most tokens are billed at the lower rate. One version is an executor that escalates hard decisions to an advisor model. Another is an orchestrator that hands bulk work to lower-cost workers. Both have to earn their place in measurement against the best single-model configuration.
Checkpoint 2 of 9· Check yourself
A team is building an agent that drafts merger-contract clauses unattended for hours. Lawyers must rewrite any clause it gets wrong, and the firm says accuracy matters far more than token spend. Which starting plan fits?
A wrong answer is expensive here, and accuracy outweighs cost, so capability-first applies. You start on Opus 5.5 and cut cost only where the evals show quality holds. Starting on Haiku means learning from rejected work. Fable 5.2 is the step to take when Opus at xhigh or max effort still falls short. A parallel majority vote multiplies cost and latency.
“you might want to start capability-first: implement with the strongest starting point for your task, then optimize to more efficient models later”Source: platform.claude.com
Checkpoint 3 of 9· Exam question
A support product routes each incoming ticket to one of eight queues. Volume is very high, the routing call must return well under a second, and a labelled sample of 500 tickets is available. Which approach to picking a tier is most appropriate?
Correct answer: C — Use the smallest tier for routing, then score it on the 500 tagged tickets and move up only if accuracy falls short.
- A. Incorrect. Adopting a middle tier and waiting for agents to report misroutes leaves the choice unmeasured before launch. A labelled sample is available, so each tier can be scored up front instead of letting misrouted tickets reach customers first.
- B. Incorrect. Calling three tiers per ticket multiplies cost at very high volume, and the response is only as fast as the slowest call. That puts the sub-second routing budget at risk, and the vote still does not say which single tier is adequate.
- C. Correct. Eight-way ticket routing under a tight latency budget and heavy volume is typically within reach of the smallest, fastest tier. Scoring it on the labelled sample checks that before launch, and a larger tier is used only if measured accuracy misses the target.
- D. Incorrect. The most capable tier is the slowest and most expensive, so using it for a very high volume of calls that must return well under a second is unlikely to fit. Tuning its prompt raises accuracy but does nothing about that latency and cost, and a simple eight-way classification rarely needs that headroom.
Checkpoint 4 of 9· Exam question
A team is building an autonomous coding agent that works for hours on large refactors. Failures are expensive because engineers must review and redo bad pull requests, and cost per run is secondary. How should they choose a tier?
Correct answer: A — Start on the most capable tier and tune prompts there, then step down in tier or effort only for tasks where evals show equal quality.
- A. Correct. Bad pull requests cost engineer time, and per-run spend is not the constraint, so the sound path is to establish the best achievable quality first and then move down a tier or effort level only where measurements show no loss.
- B. Incorrect. Fitting the tier to a spend limit puts the secondary factor first. Retries after a failed review still hand reviewers weak pull requests and add runs that the team gains nothing from.
- C. Incorrect. Planning decides how a large refactor is structured, so putting it on the fastest tier risks flawed plans that the strongest editor then carries out faithfully. The scenario's problem is output quality, not session speed.
- D. Incorrect. A small tier suits cheap, simple work. Here, waiting until most pull requests are rejected means engineers repeatedly review and redo poor changes, which is the costly outcome the team wants to avoid, and tuning prompts on a weak tier hides what stronger models could achieve.
3.Try a different effort level before changing tier
Effort trades intelligence against latency and cost within a single model, so it is the first setting to change when a tier almost meets its target. Tuning effort is often a better lever than switching models. Defaults differ by model. Fable 5.2 and Sonnet 5.5 default to high, and Opus 5.5 and Haiku 5.5 default to medium. Start at the default and adjust up or down based on your evals. How effort works is covered in generation mechanics and model options.
Which way you move depends on the problem. If cost is too high and quality is fine, try lower effort on the model you use now. If quality falls short and latency has room, raise effort and run the eval again. If the gap closes, you have paid for extra thinking on one model instead of a tier whose per-token prices are double. Opus 5.5 lists at $4 / $20 against Sonnet 5.5's $2 / $10. Escalation follows the same rule at the top of the range: Fable 5.2 is the choice when your evals on Opus 5.5 at higher effort still fall short.
Effort changes quality by a lot in both directions, so its effect has to be measured, not assumed. In Anthropic's tests on long coding tasks, medium effort cost about 30% to 52% less than high. Depending on the task and the model, lower effort also cost up to 34 points of score.
Raising max_tokens is not a fix for wrong answers. It helps only when attempts are being cut off, which shows up as responses ending with stop_reason: max_tokens. If the model returns complete answers that are wrong, more output room changes nothing.
Checkpoint 5 of 9· Match them up
Each situation below comes from a deployment's eval results. Match it to the change that fits it first.
Tap a term, then the definition that fits it.
Effort is the first lever within a model. Lower it to save cost, and raise it to close a quality gap. max_tokens matters only when output is being cut off. A tier change comes after effort can't close the gap, and Fable 5.2 is the step when Opus 5.5 at higher effort still falls short.
“Costs are too high; quality is fine | Try lower effort on the model you use now”Source: platform.claude.com
Checkpoint 6 of 9· Exam question
An analytics team's mid-tier model passes 91% of its eval set, but the target is 94%. Latency is comfortably within budget. Before moving every request to a higher tier, what should they try first?
Correct answer: B — Raise the effort setting on the current model and rerun the eval to see whether the gap closes at a smaller cost increase.
- A. Incorrect. Stripping tuned instructions removes the context that produced the current 91%, so it is more likely to lower the score than raise it.
- B. Correct. Effort trades intelligence against latency and cost inside one model, so it is a cheaper lever to test than a tier change, and the eval tells you whether it closes the three-point gap.
- C. Incorrect. A larger
max_tokensonly raises the ceiling on output length. It does not improve reasoning quality unless outputs are being truncated, which the stem does not say. - D. Incorrect. A three-point shortfall does not prove the tier is wrong. Jumping straight to the top tier adds cost and latency before cheaper levers have been measured.
4.Backing the choice with your own evals
A tier choice is justified by results on your own workload. Neither a public ranking nor a vendor's published figures can stand in for them. Build benchmark tests specific to your use case and run them with your actual prompts and data. Compare models on accuracy, response quality and handling of edge cases, then weigh the results against cost. Even Anthropic describes its own cost and quality measurements as internal and directional, not guarantees. A model that tops a general benchmark can still do worse than your current model on invoice parsing, and only an eval built from invoices will show that.
A method that holds up has four steps:
- Pull a few tasks from production logs, weighted like real traffic. Write outcome checks for each one, such as tests pass, ticket closed or row count correct, and record the cost per task beside the score.
- Baseline the tiers at several effort levels, not just the default, and plot score against spend.
- If the curve shows a gap that effort cannot close, add a multimodel strategy and run the suite again.
- Run the winning configuration in shadow on a slice of traffic before cutover, then keep the suite running.
Compare configurations on cost per completed task, not cost per token. A per-token price tells you nothing about how many tasks get solved. How to price a task from its token counts is covered in cost diagnosis, and how to build the graders is covered in evaluation design. The justification you end up with is a claim a reviewer can check: this tier, at this effort level, met the accuracy target within the latency and budget criteria, at the lowest cost per completed task of the configurations that passed.
Checkpoint 7 of 9· Put it in order
Put these steps for backing a tier choice with evals on your own workload in the right order.
- 1.Pull a few tasks from production logs, weighted like real traffic, write an outcome check for each, and record the cost per task beside the score
- 2.Baseline the tiers at several effort levels, not just the default, and plot score against spend
- 3.Run the winning configuration in shadow on a slice of traffic before cutover, then keep the suite running
- 4.If the curve shows a gap that effort cannot close, add a multimodel strategy and run the suite again
You need a task set that reflects real traffic before you can measure anything. Baselining across tiers and effort levels then shows whether a gap remains that only a multimodel design can close. The winning setup goes into shadow before cutover, and the suite keeps running afterward so it can catch changes when the model under your application changes.
“Baseline the model tiers across effort levels, not only the default, and plot score against spend.”Source: platform.claude.com
5.What changes when the model changes
A model ID names one fixed snapshot. When you use a model ID in a request, the underlying model stays the same for the lifetime of that ID. Anthropic does not update the weights or configuration of an existing ID. A new version ships under a new ID. From the 4.6 generation onward, a dateless ID such as claude-sonnet-4-6 is the snapshot itself, not a pointer that always routes to the latest version. Earlier models also have convenience aliases: claude-sonnet-4-5 resolves to the most recent dated snapshot for that minor version. The stability guarantee covers model IDs, not those aliases.
| Identifier | What it is | Can the model behind it change? |
|---|---|---|
| claude-sonnet-4-6 (dateless, 4.6 generation and later) | The canonical ID, mapped to a single fixed snapshot | No: updated versions ship under a new model ID |
| claude-sonnet-4-5-20250929 (dated, earlier models) | A pinned snapshot with the date in the ID | No |
| claude-sonnet-4-5 (alias, earlier models) | A convenience pointer to the most recent dated snapshot for that minor version | Not covered by the per-ID stability guarantee |
Behavior can still change, in three ways. The first is switching to a new ID, which changes more than the weights. Defaults such as effort move with the model. Opus 5.5 also runs with adaptive thinking always on, so audit your integration for three things before switching: requests that force tool use with a tool_choice of type any or tool, code that turns thinking off, and use of the computer_20251124 computer use tool. Each of those is a documented breaking change on Opus 5.5. The second way is retirement, which forces a switch on a schedule you do not set. The third is infrastructure. The request router, safety classifiers and sampling logic can change around a fixed ID and cause minor differences in observable behavior even when the weights have not changed. If a previously stable ID starts behaving differently, an infrastructure update is the most likely cause.
Upgrades usually help, but not evenly. In Anthropic's measurements, moving to a newer model usually cost less, up to 87% less per solved task on one coding benchmark. Yet Fable 5.2 scored lower on one research benchmark. Prompts written for the old model can also go stale: auditing prompts against the new model saved 14% on both upgrades Anthropic tested. Measured averages cannot tell you which side of that spread your workload falls on. Your own suite can.
Checkpoint 8 of 9· Check yourself
An app that writes multistep research reports runs on claude-opus-5-5. Fable 5.2 has shipped, and Anthropic's measurements show newer models usually cost less per solved task. What should the team do before moving to it?
Upgrades usually help, but not on every workload: Fable 5.2 scored lower on one research benchmark. Anthropic's averages can't show which side of that spread this app falls on, but the team's own suite can. A model ID is a fixed snapshot and never starts routing to a newer model. Default effort does not guarantee quality.
“None on coding, a gain on most; lower on one research benchmark for Claude Fable 5.2”Source: platform.claude.com
6.Planning a migration with regression evals
| State | Meaning | What to do |
|---|---|---|
| Active | Fully supported and recommended for use | Keep the regression suite running against it |
| Legacy | Will no longer receive updates and might be deprecated in the future | Start testing replacements |
| Deprecated | Still works but is no longer recommended; has a named replacement and a retirement date | Migrate all usage before the retirement date |
| Retired | No longer available; requests fail | Too late: production calls are already failing |
Anthropic gives at least 60 days' notice before retiring a publicly released model, by email and in the documentation. Treat that as a deadline, not as the time you have for testing. The guidance is to test your application with newer models well before the retirement date. Every current model already lists a retirement date it will not come before. For Claude Opus 5.5 that is not sooner than September 22, 2027, so you can plan the migration long before any notice arrives.
Check retirement dates per platform. The dates on Anthropic's deprecations page apply to the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Microsoft Foundry. Amazon Bedrock and Claude Models on Google Cloud (Google managed) set their own retirement schedules, so the same model's status and dates can differ there.
A migration plan runs in this order:
- Find where the old model is still in use. Go to the Usage page in Claude Console, click Export, and review the downloaded CSV, which breaks usage down by API key and model.
- Run the existing eval suite against the replacement and compare it with the current model on the same tasks. This A/B test against a baseline model or earlier version is your regression check, and the old model's scores are the bar to meet.
- Re-baseline effort and audit prompts on the new model, because defaults and behavior did not carry over.
- Run the replacement in shadow on a slice of traffic, cut over, and keep the suite running so that later infrastructure changes are caught too.
Checkpoint 9 of 9· Match them up
A team audits the lifecycle status of every model their application calls. Match each status to what it means for that application and what the team should do.
Tap a term, then the definition that fits it.
A deprecated model still answers requests, but it has a replacement and a retirement date, so you must migrate before that date. After retirement, requests fail. A legacy model is the earlier warning that tells you to start testing replacements.
“Deprecated: The model is still functional but no longer recommended. Anthropic provides a recommended replacement and assigns a retirement date.”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.
“Choosing a Claude model means balancing capabilities, speed, and cost.”
↩︎ Comparing tiers against hard limits“Fast mode delivers up to 2.5x higher output speed at premium pricing on supported models.”
↩︎ Comparing tiers against hard limits“Upgrade only if necessary for specific capability gaps.”
↩︎ Two ways to pick a starting tier“Consider increasing efficiency by lowering effort or downgrading models over time with greater workflow optimization.”
↩︎ Two ways to pick a starting tier“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.”
↩︎ Two ways to pick a starting tier“Several Claude models support an effort parameter that trades intelligence for latency and cost within a single model.”
↩︎ Two ways to pick a starting tier“the xhigh effort level, between high and max”
↩︎ Two ways to pick a starting tier“Multimodel strategies pair a lower-cost model with a frontier model so that most tokens are billed at the lower rate.”
↩︎ Two ways to pick a starting tier“Tuning effort is often a better lever than switching models.”
↩︎ Try a different effort level before changing tier“Create benchmark tests specific to your use case. Having a good evaluation set is the most important step in the process.”
↩︎ Backing the choice with your own evals“Test with your actual prompts and data.”
↩︎ Backing the choice with your own evals“If your integration forces tool use (tool_choice of type any or tool), turns thinking off”
↩︎ What changes when the model changes“is built for long-running agentic coding and knowledge work, with adaptive thinking always on.”
↩︎ What changes when the model changes“Create benchmark tests specific to your use case. Having a good evaluation set is the most important step in the process.”
↩︎ Key concept“On Claude Opus 5.5 and Claude Haiku 5.5, the default is medium; start there and adjust the same way.”
↩︎ Exam trap 3“you might want to start capability-first: implement with the strongest starting point for your task, then optimize to more efficient models later”
↩︎ Checkpoint“On Claude Opus 5.5 and Claude Haiku 5.5, the default is medium; start there and adjust the same way.”
↩︎ Prediction - 2.
“What is the acceptable response time for the model?”
↩︎ Comparing tiers against hard limits“Most use cases need multidimensional evaluation along several success criteria.”
↩︎ Comparing tiers against hard limits“A/B testing: Compare performance against a baseline model or earlier version.”
↩︎ Planning a migration with regression evals - 3.https://platform.claude.com/docs/en/about-claude/models/optimizing-for-cost-and-intelligenceOfficial docs
“None: twice the standard token prices”
↩︎ Comparing tiers against hard limits“A multimodel configuration must beat the single model’s whole curve.”
↩︎ Two ways to pick a starting tier“Try lower effort on the model you use now”
↩︎ Try a different effort level before changing tier“Up to 34 points lower, depending on the task and the model”
↩︎ Try a different effort level before changing tier“On long coding tasks, medium cost about 30% to 52% less than high, and low about two thirds less”
↩︎ Try a different effort level before changing tier“Attempts end with stop_reason: max_tokens”
↩︎ Try a different effort level before changing tier“These results are Anthropic-internal (Benchmarks referenced) and directional, not guarantees, so measure on your own workload with the four-step method.”
↩︎ Backing the choice with your own evals“Baseline the model tiers across effort levels, not only the default, and plot score against spend.”
↩︎ Backing the choice with your own evals“Compare on cost per completed task, not per token”
↩︎ Backing the choice with your own evals“None on coding, a gain on most; lower on one research benchmark for Claude Fable 5.2”
↩︎ What changes when the model changes“14% on both model upgrades tested”
↩︎ What changes when the model changes“Run the winner in shadow on a traffic slice before cutover, then keep the suite running.”
↩︎ Planning a migration with regression evals“Attempts end with stop_reason: max_tokens”
↩︎ Exam trap 1“Costs are too high; quality is fine | Try lower effort on the model you use now”
↩︎ Checkpoint - 4.https://platform.claude.com/docs/en/models/overviewOfficial docs
“Use Claude Fable 5.2 for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short.”
↩︎ Try a different effort level before changing tier - 5.
“When you use a model ID in an API request, the underlying model remains constant for the lifetime of that ID.”
↩︎ What changes when the model changes“An alias such as claude-sonnet-4-5 is a convenience pointer that resolves to the most recent dated snapshot for that minor version.”
↩︎ What changes when the model changes“If you notice unexpected behavioral differences on a previously stable model ID, an infrastructure update is the most likely cause.”
↩︎ What changes when the model changes“A common misconception is that dateless model IDs such as claude-sonnet-4-6 behave as evergreen pointers that route to the latest or best-performing version.”
↩︎ Exam trap 2 - 6.
“Deprecated: The model is still functional but no longer recommended. Anthropic provides a recommended replacement and assigns a retirement date.”
↩︎ Planning a migration with regression evals“providing at least 60 days’ notice before model retirement for publicly released models.”
↩︎ Planning a migration with regression evals“Partner-operated platforms, Claude Models on Google Cloud (Google managed) and Amazon Bedrock, set their own retirement schedules”
↩︎ Planning a migration with regression evals“Review the downloaded CSV to see usage broken down by API key and model.”
↩︎ Planning a migration with regression evals“Partner-operated platforms, Claude Models on Google Cloud (Google managed) and Amazon Bedrock, set their own retirement schedules”
↩︎ Exam trap 4