Business Value Pillars for Claude Solutions: Cost and Efficiency

1.Business value pillars start from cost per outcome

The exam guide names five business value pillars for a Claude solution: efficiency, transformation, productivity, cost and performance SLAs. Anthropic's documentation does not publish these as a formal framework. What it does give is a set of measured levers, and this lesson sorts those levers under the pillar each one serves. This page covers cost and efficiency. Performance SLAs, productivity and transformation each get their own treatment.

All five pillars share one starting point. A design is valued by what a finished outcome costs and delivers, not by what a token costs. Anthropic's cost guide says this outright for model decisions: when you choose or switch models, compare on cost per completed task. Anthropic's Enterprise cost guide turns the same idea into two business questions. First, what would this work have cost without AI, in resources, in time, or because the project would never have been attempted? Second, is the task hard (it needs judgment and reasoning), or just large (a high volume of straightforward work)?

The second question is where architecture begins. If you give complex reasoning to a cheaper model, the finished task often ends up costing more, because the model burns tokens on retries and needs more human correction. If you put a frontier model on basic document processing, you pay for capabilities the task never uses. So real designs often mix models. In the guide's insurance example, a frontier model helps an adjuster evaluate a complex commercial claim, while Haiku tags and triages the documents feeding into it.

Every number on this page comes from Anthropic's own runs. The cost guide calls these results directional, not guarantees, and tells you to measure on your own workload. The model-selection guide sets out the procedure for that measurement, and the order of its steps matters.

Sources12

2.Cost: use the free wins before the tradeoffs

A prototype can ignore cost, but a production system cannot. Anthropic's cost guide puts it this way: when a workload moves to production, cost becomes a first-class design constraint. The guide sorts cost levers into two kinds, and the difference between them is the main thing to learn for the cost pillar.

Free wins cut spend without touching quality. They are prompt caching, token hygiene, auditing your prompts against the model you are running, batch processing at 50% off for work that can wait up to 24 hours, and workspace spend limits as a backstop. Tradeoffs give up some intelligence to save cost. They are model choice, effort, output caps and task budgets, an elapsed-time clock, and multi-model architectures. The design rule follows: use up the free wins first, then decide how much quality you are willing to trade.

Among the tradeoffs, the model-selection guide says tuning effort is often a better lever than switching models. In Anthropic's runs, medium effort on knowledge work saved 13% to 31% and cost 1 to 3 points of quality. Effort can also differ from one call to the next. The Enterprise guide recommends low effort for routing and extraction and high effort for the final recommendation, so you pay peak rates only where they matter.

Multi-model strategies pair a lower-cost model with a frontier model so that most tokens are billed at the lower rate. The guide found that a second model paid off in two shapes: an advisor, which a cheaper executor consults on hard decisions, and an orchestrator, which hands bulk work to cheaper workers. These per-token list prices are what the model mix is spending:

Sources13

3.Efficiency: caching, batching and lean context

In this lesson, efficiency means getting the same output for fewer resources. The levers that do this are mostly the free wins.

Prompt caching comes first, because an agent loop resends its whole history on every turn. A 40-turn task sends its first turn 40 times, so the cost of a task grows with roughly the square of the number of turns. Caching does not stop the resending, but the cached prefix is billed at the cache-read rate, which is a tenth of the input price. Cache writes cost 1.25x the input price for the default 5-minute lifetime and 2x for the 1-hour lifetime. Use the 1-hour lifetime when a person pauses between turns. As a benchmark, agent loops over a day of real traffic read a median 84% of their input from the cache. Below about 80%, look for something that is breaking the cache.

Batch processing handles work nobody is waiting for. The Message Batches API processes requests asynchronously at 50% off. Most batches finish in under an hour, but results are only guaranteed within 24 hours, and a batch that does not finish in that time expires. One batch holds at most 100,000 requests or 256 MB, whichever limit is reached first. Typical uses are large-scale evaluations, content moderation and bulk content generation.

Token hygiene and prompt audits remove waste. In Anthropic's runs, tool search saved 45% when 500 tool definitions were attached. Passing data files through code execution saved 92% on a data task and also raised accuracy from 6 of 25 answers to 25 of 25. Auditing prompts against the current model saved 14% on both migrations measured, and ran faster because it needed fewer tool rounds.

Sources14

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.The model with the lowest price per token always gives the cheapest solution.Why is that wrong?

   The measure that matters is cost per completed task. A cheaper model on hard reasoning can cost more in total because of retries and human correction.

   Covered in Business value pillars start from cost per outcome

2. 2.Batch processing is a free 50% discount, so every workload should use it.Why is that wrong?

   The batch discount costs you latency. Results can take up to 24 hours, so batching only fits work that nobody is waiting on.

   Covered in Efficiency: caching, batching and lean context