Spring AI Token Usage: Measure Cost Before You Pick a Model — LLM Cost Control 1/4

Short summary

This article is the first in a four-part series on controlling LLM costs in Spring AI, covering observability setup and model selection as foundational cost drivers. It explains how to use Spring Boot's Micrometer-based observability to attribute token usage to specific application features, and recommends creating per-cost-tier ChatClient beans so tasks route to the cheapest capable model. The author stresses treating observability as step zero before any optimization, enabling measurable before-and-after comparisons for every subsequent change.

• •Set up Micrometer observability with custom low-cardinality tags to attribute token usage to specific features before optimizing anything
• •Route tasks to the smallest capable model using per-cost-tier ChatClient beans (e.g. gpt-5-mini for classification, gpt-5 for complex analysis)
• •Configure alerts for sudden token-usage spikes and provider-side spending limits as final safeguards against runaway costs

Generated with AI, which can make mistakes.