LLM Cost Reduction
Your LLM bill grows faster than your traffic. In 2–3 weeks I work in your code and bring the cost per task down without losing quality: every change passes a quality gate on an eval set built from your real traffic, ships behind a flag and can be rolled back.
Who it’s for: For companies with LLMs in production — product features, agents, internal pipelines — paying a real API bill every month and watching it grow.
Not for teams before production — that’s where the Opportunity Map starts. Broken builds go to the Rescue. If the question is “which model for which task”, the AI Usage Audit answers it.
The bill grows faster than the traffic.
The first deployment put the strongest model on everything, because back then the point was to make it work. Since then you’ve added agents with retry loops, long contexts and a RAG pipeline that puts more into the prompt than the task needs — and the API invoice climbs every month while traffic doesn’t. The team is afraid to touch the setup because nobody can say whether a cheaper variant would hurt quality. That is measurable: first an eval set from real traffic and a cost per task, then changes that pass the quality gate.
What you get
- An eval set built from your real traffic, with baseline quality and cost per task for every call site
- A quality gate: a change ships only if it scores at least as well as today on that set
- Changes implemented in your code, behind a flag and on shadow traffic, with rollback: routing to cheaper models with escalation, prompt and response caching, shorter context, output length limits, batch for jobs that can wait
- A smaller or fine-tuned model where the task is narrow and the volume is high
- Before/after on the same set: cost per task, quality, latency
- Cost and quality monitoring with alerts, plus written rules so the bill doesn’t creep back
How it runs
What you need on your side
This works if you have:
Price
from €2,000net · measurement
Quoted after the measurement (days 1–3), always before the first change ships, in one of two billing variants. Both carry the same guarantee: if after the implementation the cost per task on the same eval set hasn’t dropped by at least 20% at quality no worse than today, you pay only for the measurement (days 1–3). The changes stage is then on me.
- capped at 2–3 weeks — I don’t move in for months
- if the measurement itself shows less than a 20% cut in cost per task is available, I say so and we stop there — you keep the eval set
Tell me your monthly model API bill and what the models do. Within 1–2 business days I’ll reply with the measurement price and a start date.
Or by email: michal@jaskol.ski
As of: November 2026
Run your own numbers.
Take your monthly bill for LLM APIs and a conservative share that can come off at the same quality. What stays in the company over a year?
The math
less spent on LLMs over 12 months, at the same quality
€36,000 — less spent on LLMs over 12 months, at the same quality- less on the LLM bill per month
- €3,000
- total over 12 months
- €36,000
An estimate built on the numbers you enter. The real share comes out of the measurement in days 1–3 — if it’s under 20%, we stop at the measurement.
Compare it with the price of the changes: email me your monthly API bill.
Before you write.
How do we know quality didn’t drop?
From the quality gate. In the first three days we build an eval set from your real traffic and measure today’s quality on it. A change ships only if it scores at least as well on that same set, and it goes to production behind a flag or on shadow traffic, with rollback. At the end you get a before/after on the same set: cost per task, quality and latency.
Isn’t this just switching to a cheaper model?
A model swap is one lever, and rarely the best one. Routing usually gives more: a cheaper model for the easy cases, escalation to the strong one only where it’s needed. Then prompt and response caching, shorter context and fewer RAG chunks, output length limits and structured outputs, batch for jobs that can wait, and cutting retry loops in agents. Which lever to pull first comes out of the measurement in days 1–3.
How much can we save?
I don’t know until I measure, and I wouldn’t trust anyone who quotes a percentage before measuring. It depends on what share of your traffic is easy cases, how much context goes into every call and how many repeats can be served from cache. That is why the first three days are a measurement with an estimate per lever, and the quote for the rest comes after it. If there is less than 20% to gain, I say so and we stop at the measurement.
What if the saving comes out smaller?
There are two thresholds. The first after the measurement: if the levers add up to less than 20%, I say so and we stop at the measurement. The second after the implementation: if the cost per task on the same eval set hasn’t dropped by at least 20% at quality no worse than today, you pay only for the measurement and the changes stage is on me. At the quote you also pick how to pay: a fixed price for a fixed scope, or a lower fixed price plus 20% of the measured monthly saving for 6 months. The saving is computed in the open: (cost per task before − cost after) × number of tasks in the month, from the monitoring that stays with you.
Will we end up locked into one vendor?
The opposite. Routing and the quality gate work with any model, so switching vendors becomes a configuration change you can measure rather than a rewrite. Where the task is narrow and the volume is high, the candidate is a smaller open-weight or fine-tuned model — hosted by you or by a provider, behind the same gate.
How is this different from the AI Usage Audit?
The audit asks whether you use AI where you should and in the best available setup, and ends with a ranked list of upgrades. Here I work in the code with one goal: the same result for less. If you don’t know whether the problem is cost or the choice of tools, start with the audit: cost is one of its inputs.
What’s next
Your team keeps the eval set, the quality gate and the monitoring — every future model change goes through the same check. The natural next step is an Agentic Workflow Sprint: the same measurement, the next process.