Why A is the answer

Route by complexity, with a fast tier by default and a stronger tier for hard cases

Routing puts spending where it changes the result. One tier for everything either overpays on simple work or fails on hard work, and asking people to choose on each reply is not a strategy.

In detail

• Traffic varies in difficulty, so one tier cannot suit it all.
• Routing spends money where it changes the result: a fast tier by default, and a stronger one for hard cases.
• Measure quality on each route, because a router can send hard cases the wrong way.

See it in different ways

flowchart LR
  A["Request"] --> B{"Complexity check"}
  B -->|"Simple"| C["Fast model"]
  B -->|"Hard"| D["Most capable model"]
  C --> E{"Validation passes?"}
  E -->|No| D
  E -->|Yes| F["Answer"]

sequenceDiagram
  participant U as Request
  participant F as Fast model
  participant V as Validation
  participant S as Stronger model
  U->>F: Handle first
  F->>V: Check the output
  alt Fails validation
    V->>S: Escalate
  else Passes
    V-->>U: Return the answer
  end

flowchart TD
  Q{"How hard is the ticket?"}
  Q -->|"Simple"| A0["Fast tier"]
  Q -->|"Needs deep reasoning"| A1["Stronger tier"]

Why the other options are wrong

B. Let each support agent choose a model from a dropdown on every reply

Asking people to choose for each reply is slow and inconsistent.

C. Use the largest model everywhere, so quality never becomes a question

This overpays on simple tickets, and every reply gets slower.

D. Use the smallest model everywhere and let staff fix the harder replies

This fails on the hard cases and moves the cost to staff.

Related questions

• Question 20A high-volume classifier sometimes needs deeper judgment. Which design balances cost and quality?
• Question 24A latency target is hard to hit, and answers are long. What do you try first?
• Question 22A long-running agent loses track of its goal after many turns. What is the best context strategy?

Learn more

• Our lesson: Models, prompts and context
• Models overview
• Prompt caching
• Effective context engineering