This sounds more like a semantic caching layer on top of OpenRouter. The real challenge is knowing which parts of previous work are still reusable and only sending the delta to the model.
We process a lot of business documents at scale, invoices, sales orders, insurance documents and so on. Our token usage easily go beyond 50 to 100 million tokens per day.
We use a routing approach where simpler tasks go to smaller, cheaper models and more complex tasks go to frontier models. We are also working on having our own model for some of these workloads.
Document processing is probably one of our biggest agent use cases for tokens at this scale.
Is this through LLM providers on a metered API connection?
You're probably at the scale that having your own hardware to run local models is worth it, even if it's a couple of mac minis or mac pros in a cupboard in your office.
