LLM cost optimisation: model routing, caching and a private model where APIs get expensive
We find where your LLM money goes and move each request type to the cheapest model that passes your quality tests. Nothing changes without a measurement.
What hurts
The API bill grows faster than revenue every month. Your CFO asks what exactly you are paying for, and there is no clear answer.
Every request goes to the most expensive model because that was simpler at the start. Yet most of them are simple: classification, short replies, field extraction.
You are afraid to cut costs because last time a cheaper model started making mistakes and customers complained.
What we build
- Cost audit by request type, agent and customer
- A quality test set for each request type
- Routing: simple requests to cheap models, hard ones to strong models
- Caching of repeated requests and provider prompt caching
- Shorter prompts and context without quality loss
- Break-even analysis for a self-hosted model on vLLM
- A cost dashboard with breakdowns and overspend alerts
What you get
- A cost report showing where the money goes
- A model router and cache configured in your code
- A test set that protects quality from cost cutting
- A cost dashboard with budget alerts
- A plan of next steps with payback estimates
How we measure the result
- Cost per conversation or document before and after
- Monthly LLM bill at the same volume
- Test set score after the changes
- Share of requests served by a cheaper model or the cache
Systems and stack
Industries
Questions
How much does LLM cost optimisation cost?
From $5,000 for a 3–5 week project. The price depends on the number of AI scenarios and whether requests are already logged. It is worth doing when your monthly API bill is visible in the budget, otherwise savings will not cover the work.
Will quality drop?
That is the main risk, so we build the test set first and only then change models. Each change passes the evals and ships to part of the traffic. If quality drops, we restore the previous setup.
How much can we save?
We do not promise numbers upfront because they depend on your requests. After the cost audit we show a calculation per scenario and you decide what we do. The savings metric can be written into the contract.
Which providers do you work with?
OpenAI, Anthropic, Google Gemini, Mistral and open-weight models on vLLM. Our Happ platform also runs on several models plus our own LLM for call analysis, so we do this for ourselves too.
What do you need from us?
Access to provider billing, request logs and the code that calls the models. If there are no logs, we switch them on for a week first to see the real load.
When is a self-hosted model cheaper than an API?
When requests are numerous, similar to each other and spread evenly across the day. Then a GPU server pays off, and your data stays with you as a bonus. With small or uneven load an API is usually cheaper. We calculate both options on your logs, including the cost of running the server.
Related services
More in this section
Describe the task. We reply within one business day, the first call is with the founder.
Describe the task. The first call is with the founder.
We reply within one business day. 30 minutes to tell you whether it is worth doing and what it costs.