Spring AI Tips to Cut LLM Costs by Managing Output Tokens and Chat Memory

A developer guide published on DEV Community outlines practical strategies for controlling large language model (LLM) costs using Spring AI, focusing on output tokens, reasoning tokens, and conversation history. Output tokens typically cost five to eight times more than input tokens, and reasoning models compound this by generating hidden thinking tokens billed at the same higher rate. Spring AI offers a provider-independent maxTokens cap and provider-specific reasoning controls, such as OpenAI's reasoning-effort setting and Anthropic's thinking-token budget, to limit runaway response costs. The guide warns that a Spring AI 2.0 upgrade raised Anthropic's default maxTokens from 500 to 4096, potentially multiplying response costs for users who relied on the old cap without realising it. Developers are also cautioned to disable reasoning during local Ollama development, since models like qwen3 and deepseek-r1 use reasoning by default, making hidden token usage easy to overlook before moving to a billed production environment.