Practical cost optimization for LLM-heavy stacks

What tactics moved the needle? - Prompt slimming: Remove boilerplate and unused context; enforce hard caps. - Caching: Hash prompts; share caches across services. - Routing: Use cheaper models for easy tasks; escalate on uncertainty. - Compression: Summarize long threads before passing to models. - Batching: Group small requests to hit better throughput on inference servers. - Observability: Track cost per endpoint and per customer. Share your ROI stories and any gotchas when downgrading models.