Token optimisation: the same quality — a fraction of the price
Everything an AI reads and writes costs tokens — and tokens cost money. Token optimisation is one of my very biggest focus areas: I design solutions where costs do not explode as usage grows. Smart AI is only half the job; cheap operation is the other half.
- The most expensive model is used for everything — including tasks a small model handles just as well.
- Entire documents and long histories are sent along with every single call, again and again.
- Nobody knows what a single task costs — the bill is a black box that grows with usage.
- What was cheap in the pilot phase becomes an unpleasant surprise in production.
- Every task hits the cheapest model that can solve it — the big models are saved for the hard ones.
- Repeated content is cached, and only the context that is actually relevant is sent along.
- Usage is measured per task, so you know the price of an invoice, an email or a set of minutes.
- Costs are predictable and scale gracefully when usage grows tenfold.
How I optimise
Measure before optimising
I start by measuring token usage per task, so I know exactly where the money is going — and where the gain is biggest.
The right model for the task
Simple tasks hit small, fast models; only the hard ones go to the big models. That routing alone often cuts the bulk of the bill.
Prompt caching
Instructions, knowledge bases and examples that recur in every call are cached — so you do not pay full price for the same content thousands of times.
Only relevant context
Instead of sending the entire document archive along, the solution retrieves only the passages relevant to the task at hand.
Sharp, structured output
The model is asked for exactly the output that is needed — not three paragraphs of polite wrapping. Shorter answers are both cheaper and easier to work with downstream.
Continuous monitoring
Usage is monitored in production with alerts for anomalies — so a bug or a new usage pattern never gets the chance to become an expensive surprise.
Typical tools in the toolbox
The techniques are the same whether the solution runs on OpenAI, Anthropic, Google or a European provider — I pick what fits the task and the budget.
What it typically means
AI running costs once routing, caching and context management are in place
the optimisation happens behind the scenes — answers do not get worse, and are often faster
You know the price per task and can scale usage without fearing the bill
The saving depends on how unoptimised the starting point is — I always measure before and after, so the gain is there in black and white.