What you will be able to do
- Define a value-based architectural decision as the option with the largest measured gain on stated value pillars that stays within a cost or SLA ceiling
- Turn value pillars such as efficiency, productivity, solution cost and performance SLAs into measured baselines that competing proposals can be scored against
- Pick between single-tier, routed and multimodel designs using cost per completed task instead of cost per token
- Tell levers that cut cost without lowering quality apart from levers that trade cost for intelligence or latency
Key concept
Value-based architectural decision — Out of the designs that stay within the stated cost and SLA limits, choose the one that gives the largest measured improvement on the stakeholders' named value pillars. The ceiling is a fixed constraint. Capability, cost and component count matter only through what they do to the measured pillars.
1.What makes an architectural decision value-based
A value-based decision starts from two inputs the business has already given you: the value pillars it cares about, and a ceiling it will not cross. The usual pillars are efficiency, transformation, productivity, solution cost and performance SLAs. The ceiling is often stated as cost per unit of work or as a response-time commitment. The architect's job is to find the option that stays under the ceiling and moves the named pillars furthest. The ceiling is a constraint to design within. It is not an opening bid to renegotiate once the bill arrives.
Neither end of the model range wins by default. The most capable model can cost too much at the volume the business runs, and the cheapest model can miss the quality target. Claude gives you a separate choice of model, effort level and architecture for each request, so most decisions place a workload somewhere between those two ends. They are not a pick of one end or the other. Effort is the setting that moves a workload within a single model: it is a per-request parameter that trades intelligence for latency and cost, so raising it buys more intelligence at higher latency and cost, and lowering it does the reverse.
Three shortcuts look like value reasoning but are not.
- Maximum capability. "Use the biggest tier because capability always pays" ignores the ceiling and the fact that most work does not need the top tier.
- Minimum spend. "Use whatever has the lowest token spend because cost is the only measurable pillar" is wrong on both counts. Latency, accuracy and completion rates are all measurable, and most use cases need evaluation along several criteria at once.
- Maximum components. "Add agents and tools because a richer architecture shows transformation" gets it backwards. Agentic systems usually trade latency and cost for better task performance, and that trade has to earn its place.
Checkpoint 1 of 8· Check yourself
A support organization tells you that its value pillars are agent productivity and a response-time SLA, and that each ticket may cost no more than a fixed per-ticket amount. Which approach to choosing the Claude design is value-based?
A value-based decision treats the ceiling as a fixed constraint. Within it, the decision picks the design that moves the named pillars furthest, measured on several criteria. Neither end of the model range wins by default, and the number of agents is not evidence of transformation.
“The most capable model can be too expensive at scale, and the least expensive model can fall short on quality.”Source: platform.claude.com
Checkpoint 2 of 8· Exam question
Which statement best describes a value-based architectural decision for a Claude solution?
Correct answer: A — Choosing the option that gives the greatest measurable improvement on the stakeholders' stated value pillars while staying within the stated cost and SLA limits.
- A. Correct: a value-based decision ties the choice to measurable outcomes on the named pillars and treats the cost and SLA ceilings as constraints rather than variables.
- B. Incorrect: the most capable tier can exceed the cost ceiling and is unnecessary for easier tasks, so capability alone does not equal value.
- C. Incorrect: lowest spend ignores outcomes such as quality and productivity; value is the outcome achieved under the ceiling, not the lowest cost.
- D. Incorrect: more components add cost and complexity, and complexity is not a measured outcome on any pillar; the simplest sufficient design is usually better.
2.Turning value pillars into measured baselines
A pillar can only decide between proposals once it is a number with a baseline. "Improve efficiency" can't rank anything. "Cut handling minutes per claim from today's measured figure" can. Good success criteria are specific and measurable, and they are anchored to something real. Anthropic's model example sets an F1 target that is a 5% improvement over the current baseline. That baseline is what lets you compare a proposal's projected gain fairly against another proposal's.
| Value pillar | Measurable form | Example criterion |
|---|---|---|
| Performance SLA | Operational metric: response time (ms), uptime (%) | 95% response time < 200ms |
| Solution cost | Budget for running the model: cost per API call, model size, frequency of usage | Cost per completed task under the finance cap |
| Productivity / efficiency | Task completion rates; A/B test against a baseline model or earlier version | Improvement over the current baseline on a held-out set |
| Quality (gate for every pillar) | Task-specific accuracy, edge-case handling | F1 of at least 0.85 on a held-out test set |
When proposals compete, score each one on several dimensions at once. One good criteria set for a single feature combines accuracy, non-toxic output rate, error severity and a latency percentile. A feature that raises accuracy but breaks the latency SLA has not delivered value. Neither has one that hits the SLA by giving up accuracy.
Projected gains become evidence only after a prototype has been measured. Build an evaluation set from your own use case, run the top candidates on your actual prompts and data, and record quality, cost per unit of work and latency. Then weigh performance against cost. Anthropic calls a good evaluation set the most important step in this process. A ranking built this way holds up when someone asks why. A ranking built on whose idea it was, or on which model shipped most recently, does not.
Measure today's baseline on a representative sample. Without a current figure, a 30% claim can't be compared with another proposal's gain, and you can't verify it after launch.
Checkpoint 3 of 8· Match them up
An architect has to turn each stakeholder value pillar into a measurable criterion before ranking proposals. Match each pillar to the criterion that measures it.
Tap a term, then the definition that fits it.
A pillar can rank proposals only once it is a number with a baseline. SLAs map to operational metrics such as latency percentiles. Cost maps to the budget per unit of work. Productivity maps to completion rates compared against a baseline. Quality is the task-specific accuracy gate that every proposal has to pass.
“Operational: Response time (ms), uptime (%) Quantitative methods: A/B testing: Compare performance against a baseline model or earlier version.”Source: platform.claude.com
3.Choosing an architecture under a cost ceiling
Compare options on cost per completed task, not cost per token. Finance usually states its ceiling the same way, as cost per resolved ticket or per processed claim. A cheap answer that fails has to be redone, escalated or handled by a person. Every failure therefore raises the real unit cost, even when the per-token price looks low.
The price gap between tiers is large, so where the tokens go decides the bill. At list prices, Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens, for prompts up to 100,000 tokens. Claude Opus 5.5 costs $4 and $20, which is 40 times as much on both. Take a hypothetical ticket of 5,000 input and 500 output tokens. It costs about $0.00075 on Haiku 5.5 and about $0.03 on Opus 5.5. If 85% of tickets stay on Haiku and 15% escalate, the blended cost is about $0.005 per ticket. That is roughly a sixth of an all-Opus design, and it still meets the accuracy target on disputes.
| Model | Input | Output |
|---|---|---|
| Claude Haiku 5.5 (prompts up to 100,000 tokens) | $0.10 / MTok | $0.50 / MTok |
| Claude Sonnet 5.5 | $2 / MTok | $10 / MTok |
| Claude Opus 5.5 | $4 / MTok | $20 / MTok |
| Claude Fable 5.2 | $8 / MTok | $40 / MTok |
A multimodel design follows from these numbers. It pairs a lower-cost model with a frontier model so that most tokens are billed at the lower rate. Two shapes are common:
- Executor and advisor. An executor handles the work and escalates hard decisions to an advisor.
- Orchestrator and workers. An orchestrator hands bulk work to lower-cost workers.
The escalation rule matters as much as the pattern. In Anthropic's claims-processing cookbook, the cheaper model's remaining misses were borderline cases it decided on its own without asking. Escalation should therefore trigger on low confidence as well as on known-hard categories.
The rejected designs each fail in a different way.
- Largest tier for everything. It breaks the cap now and pushes the problem back to finance.
- Retrying the small tier. Repeated attempts at the same tier do not close a measured capability gap. The rule is to upgrade only where a specific capability gap shows up. Re-running failures at higher effort only pays where a verifier can check the outputs. A verifier is an automatic check on whether an answer is correct, such as a test suite. It tells you which attempts failed and need a re-run. Without one, every output looks the same, so you cannot tell which retry worked and would have to pay for every attempt.
- An evaluator-optimizer loop on every ticket. It adds a draft-and-critique cycle, with its latency and cost, to routine tickets that already pass. Complexity is only worth its cost where it buys performance the task needs.
Checkpoint 4 of 8· Check yourself
Your offline eval shows the smallest tier passes the accuracy target on routine tickets but misses it on billing disputes, even after you have tuned the prompt. Finance caps cost per resolved ticket. A teammate wants to keep the small tier on disputes and just retry it when an answer fails. What should you do instead?
Retrying the same tier, even in a self-critique loop, does not close a capability gap you have measured. Failed answers still have to be redone or escalated, so the cost per resolved ticket keeps rising. The right move is to upgrade only the segment where the gap shows up, keep routine tickets on the cheap tier, and stay under the cap without renegotiating it.
“Upgrade only if necessary for specific capability gaps.”Source: platform.claude.com
Checkpoint 5 of 8· Exam question
You are the architect for a retailer's support-automation program built on the Claude Messages API. Finance caps spend at a fixed cost per resolved ticket. An offline eval on representative tickets shows the smallest Claude tier meets the accuracy target on routine order-status and returns tickets but misses it on billing disputes, which are a minority of volume. Which design delivers the most business value within the cost ceiling?
Correct answer: A — Route routine tickets to the smallest tier and escalate billing disputes and low-confidence cases to a larger tier, keeping blended cost under the cap.
- A. Correct: tiered routing spends the larger model only where the eval shows it is needed, so accuracy holds on hard cases while the blended cost per ticket stays under the stated ceiling.
- B. Incorrect: using the largest tier for every ticket maximizes quality but ignores the stated cost ceiling, and the plan depends on finance relaxing a constraint the architect was given.
- C. Incorrect: the eval already shows the smallest tier misses the accuracy target on disputes, and repeated retries add cost without fixing a capability gap, so the accuracy outcome is still unmet.
- D. Incorrect: a draft-and-critique loop multiplies calls per ticket, including the routine ones that never needed it, which raises cost per ticket and risks breaching the ceiling.
4.Separating free wins from tradeoffs
Cost levers come in two kinds, and a value-based design uses the free kind first. Prompt caching, token hygiene, a prompt audit and batch processing lower spend without lowering quality. Model choice, effort, output caps and multimodel architectures trade cost for intelligence. Cutting cost with a free lever leaves room under the ceiling that can then go to a stronger model where the pillars need it. The mechanics belong to cache breakpoints and cost and latency optimization. What matters here is which pillar each lever spends.
Some free wins are free only for cost. The Message Batches API charges 50% of standard prices, but it gives up latency: most batches finish within an hour, and results can take up to 24 hours. That fits a nightly evaluation run or bulk summarization. It fails a live support channel with a sub-second response SLA. Before you count a batch discount toward the solution-cost pillar, check the latency SLA.
| Lever | Kind | What it costs you |
|---|---|---|
| Prompt caching | Free win | Nothing in quality; turn it on first |
| Batch processing (50% off) | Free win on cost | Latency: work must wait up to 24 hours |
| Lower effort on current model | Tradeoff | Some intelligence; restore it if quality drops |
| Next tier up / multimodel | Tradeoff | Higher unit cost on the tokens it handles |
Published lever results are directional, so your own workload decides. On Anthropic's coding benchmarks, each newer model solved at least as many tasks as the one before it. That is a good reason to test an upgrade. It does not guarantee that the newest release gives the best outcome per dollar on your task. Anthropic labels these results internal and directional, not guarantees, and tells you to measure on your own workload.
Checkpoint 6 of 8· Match them up
Before spending headroom under a cost ceiling on a stronger model, you sort the available cost levers by what each one gives up. Match each lever to what it costs you.
Tap a term, then the definition that fits it.
Free wins such as caching cut spend without hurting quality. Batch processing is free only on cost because it gives up latency. Effort, model choice and multimodel designs are tradeoffs that exchange cost for intelligence.
“Tradeoffs exchange cost for intelligence: model choice, effort, output caps and session budgets, an elapsed-time clock, and multimodel architectures.”Source: platform.claude.com
Checkpoint 7 of 8· Check yourself
Three feature proposals are competing for the same budget. One is built on the newest Claude release, and its sponsor cites Anthropic's coding benchmarks, where each newer model solved at least as many tasks as the one before it. How should this proposal be ranked?
Anthropic labels its lever results internal and directional, not guarantees. A newer model is worth testing, but proposals are ranked by measurements on your own workload, not by which release is newest.
“These results are Anthropic-internal (Benchmarks referenced) and directional, not guarantees, so measure on your own workload with the four-step method.”Source: platform.claude.com
Checkpoint 8 of 8· Exam question
A product team has four competing feature proposals for a Claude-powered claims assistant, and leadership has named its value pillars: efficiency, productivity, solution cost and performance SLAs. You are the architect asked to recommend what to build first. Choose two actions that make the prioritization defensible.(Select 2)
Correct answers: A, B — Score each proposal against the named pillars using a baseline measured today, such as handling minutes per claim, so projected gains compare fairly.; Prototype the top candidates on a representative claim sample and record quality, cost per claim and latency, so the ranking rests on measured results.
- A. Correct: anchoring each proposal to a measured baseline on the stated pillars makes the projected gains comparable and ties the ranking to what leadership said it values.
- B. Correct: a small prototype on representative claims replaces assumptions with measured quality, cost and latency, which is the evidence needed to compare options under the pillars.
- C. Incorrect: autonomy is a design property, not a business outcome; a more autonomous feature can raise cost and risk while doing nothing for the named pillars.
- D. Incorrect: sponsor seniority is not one of the stated value pillars, so ranking by it replaces the measurable-outcome criterion with organizational politics.
- E. Incorrect: a newer model is not automatically the best value; whether it improves outcome per unit of cost has to be shown for the specific task rather than assumed.
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.”
↩︎ What makes an architectural decision value-based“Compare on cost per completed task, not per token”
↩︎ Choosing an architecture under a cost ceiling“Run at low effort and re-run failures at high”
↩︎ Choosing an architecture under a cost ceiling“You can check outputs (tests, a verifier)”
↩︎ Choosing an architecture under a cost ceiling“Free wins cut spend without touching quality”
↩︎ Separating free wins from tradeoffs“batch processing trades latency for its discount”
↩︎ Separating free wins from tradeoffs“on the coding benchmarks in that section each newer model solved at least as many tasks as the one before it”
↩︎ Separating free wins from tradeoffs“You choose the model, the effort level, and the architecture for each request”
↩︎ 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 completed task, not per token”
↩︎ Exam trap 3“These results are Anthropic-internal (Benchmarks referenced) and directional, not guarantees, so measure on your own workload with the four-step method.”
↩︎ Exam trap 4“Tradeoffs exchange cost for intelligence: model choice, effort, output caps and session budgets, an elapsed-time clock, and multimodel architectures.”
↩︎ Checkpoint“These results are Anthropic-internal (Benchmarks referenced) and directional, not guarantees, so measure on your own workload with the four-step method.”
↩︎ Checkpoint - 2.
“Several Claude models support an effort parameter that trades intelligence for latency and cost within a single model.”
↩︎ What makes an architectural decision value-based“Test with your actual prompts and data.”
↩︎ Turning value pillars into measured baselines“Having a good evaluation set is the most important step in the process.”
↩︎ Turning value pillars into measured baselines“Weigh performance and cost tradeoffs.”
↩︎ Turning value pillars into measured baselines“Multimodel strategies pair a lower-cost model with a frontier model so that most tokens are billed at the lower rate.”
↩︎ Choosing an architecture under a cost ceiling“an executor that escalates hard decisions to an advisor”
↩︎ Choosing an architecture under a cost ceiling“Upgrade only if necessary for specific capability gaps.”
↩︎ Choosing an architecture under a cost ceiling - 3.
“Most use cases need multidimensional evaluation along several success criteria.”
↩︎ What makes an architectural decision value-based“which is a 5% improvement over the current baseline (Achievable).”
↩︎ Turning value pillars into measured baselines“95% response time < 200ms”
↩︎ Turning value pillars into measured baselines“Most use cases need multidimensional evaluation along several success criteria.”
↩︎ Exam trap 2“Operational: Response time (ms), uptime (%) Quantitative methods: A/B testing: Compare performance against a baseline model or earlier version.”
↩︎ Checkpoint - 4.https://www.anthropic.com/engineering/building-effective-agentsSecondary source
“we recommend finding the simplest solution possible, and only increasing complexity when needed.”
↩︎ What makes an architectural decision value-based“Agentic systems often trade latency and cost for better task performance, and you should consider when this tradeoff makes sense.”
↩︎ What makes an architectural decision value-based - 5.
“a single tier either over-pays on the easy ones or under-delivers on the hard ones.”
↩︎ Choosing an architecture under a cost ceiling“The misses are the borderline escalation cases the driver decides on its own without asking.”
↩︎ Choosing an architecture under a cost ceiling - 6.
“This approach is well-suited to tasks that do not require immediate responses”
↩︎ Separating free wins from tradeoffs