AI-103 Practice Question 23: Model Router Cost
Microsoft Azure AI Apps and Agents Developer Associate (AI-103) · Test 3 · question 3 of 10
A support assistant handles a lopsided mix of traffic: most turns are one-line FAQ lookups, and perhaps one in six is a genuinely hard multi-step troubleshooting question. Every request currently goes to the most capable model in the catalogue, and the monthly bill reflects that. The team will not accept a quality drop on the hard questions, and does not want to write and maintain classification logic in the application.
What should they deploy?
Show answer and explanation
Answer:
Model router is itself a trained model that analyses each prompt in real time (complexity, reasoning demand, task type), and routes it to the most suitable underlying model, all behind a single deployment. Cheap models take the easy turns, frontier models take the hard ones, and reasoning models are used only when reasoning is required. The default Balanced routing mode picks the most cost-effective model within a small quality band; Cost and Quality modes shift that trade-off, and a model subset constrains which models may be used at all.
- B is wrong because that is precisely the routing logic the team said it would not build and maintain. Chat models also do not return a calibrated confidence to branch on, and a retry pays for two calls instead of one. - C is wrong because Global Batch is asynchronous with a 24-hour turnaround. A support assistant has a person waiting, so the discount is unusable here. - D is wrong because capping output length and fixing temperature change the *shape* of answers, not the per-token price of the expensive model, and truncating the hard answers is the quality drop they ruled out.
Reference: Model router for Microsoft Foundry concepts