AI bills grow quietly — a pilot becomes a product, tokens multiply, and finance asks hard questions. The good news: LLM spend usually has 40–70% of avoidable cost hiding in plain sight.
The four levers
Routing: send easy requests to small models and hard ones to frontier models. Caching: identical and near-identical prompts should never hit the API twice. Prompt hygiene: trim system prompts and retrieved context; tokens are the meter. Batching: non-urgent workloads run at discounted batch rates.
Measure before optimizing
Tag every call with feature, customer, and model. Cost per successful task — not cost per token — is the metric leadership should watch.
Result
In recent engagements, routing plus caching alone paid for the entire consulting project within a quarter. Datastrel implements cost observability and optimization as part of every AI build.