AI cost optimization: how to cut your OpenAI bill 60% without losing quality
Prompt caching, model routing, structured outputs and batching. Four levers that reliably cut an AI bill by half or more, with a worked example.
An AI bill that doubled in three months is usually a routing problem, not a usage problem. Here are the four levers that move the needle, in order of effort-to-reward.
1. Prompt caching
If your system prompt is over 1k tokens and you call the model more than 50 times per hour, enable prompt caching. Anthropic and OpenAI both cache prefix tokens at a 90% discount, on a chat-heavy stack this lever alone commonly removes a third of the bill.
2. Model routing by job
Not every request needs the flagship model. Route extraction and classification to Haiku 4.5 or gpt-4o-mini, route reasoning to Sonnet 4.6 or GPT-5.5. Our router averages 4-to-1 cheaper calls without quality regression on the cheap-model jobs.
3. Structured output instead of free text
When you need 12 fields, ask for 12 fields with response_format=json_schema. Free-text responses with parsing burn 3-5x more tokens to get the same data and the parser breaks weekly.
4. Batch what is not realtime
OpenAI batch API is 50% off list and runs within 24 hours. Lead enrichment, content drafts, weekly classification jobs all belong here.
A worked example
Take a stack running $7,000/mo. Prompt caching and job-level model routing typically remove around half of the flagship-model spend; structured output and batching trim most of the rest, landing near $3,000/mo with quality checks unchanged. Implementation is usually one to two engineering weeks.
If your AI bill grew faster than your usage, our AI integration audit finds the leaks. Send us your last 3 invoices for a 30-minute review.
