AI FinOps at its best. This is how you get value for money from AI.
- I know we've all seen image a hundred times. - I also know LinkedIn isn't the natural home for technical content. But today I want to break these rules - I just published my first dev.to article, and I wrote it after solving a problem I actually hit. Quick story: I was experimenting with a OpenClaw AI Assistant, topped up my Anthropic account with 20 EUR to give it a try. Couple of prompts later, the balance was gone. That moment forced me to think how to optimize it: if every request to the same frontier model - cash burns. "Write a script on TypeScript" and "Hi" pay the same price. And the trivial prompts dominate real usage. So I built a Cloudflare Worker that classifies each prompt on the fly and routes it to the right model: → Coding questions → Claude Sonnet 4.6 → Everything else → a cheap open model on Cloudflare Workers AI (Kimi-K2) ~50 lines of TypeScript. Unified logs in Cloudflare AI Gateway. The client only sees one endpoint - routing is dynamic and invisible. On a realistic 70/30 traffic mix, this cuts LLM spend by ~60-70% versus sending everything to Claude - without dropping quality on the prompts that actually need it. Full write-up, code, and setup steps in the comments 👇 What's the fastest you've ever burned through a paid API top-up? #AI #LLM #CloudflareDevelopers #EdgeComputing #DeveloperExperience
