Olypsis / Cost engineering
Lower your LLM costs.
Keep the work useful.
We find where your AI agent spends too much, test changes on real tasks, and put the fixes into the system you already run.
Discuss your agent costs- Input & context
- Model call
- Tool result
- Next model call
- Accepted result
When to think
about this
Your agent works, but its costs need a closer look.
You have an AI workflow in use, and need to understand why it costs what it does—or whether the next increase in usage makes economic sense.
- A useful answer takes several paid calls.
- Some tasks cost far more than the average.
- Tool responses or conversation history keep growing.
- A cheaper model creates more retries or review.
Our work for
Thesis Finance
We found repeated context and oversized options-tool output. We trimmed the payloads, improved caching, and tested the changes on captured options turns.
Cost per turn
57% lower replay cost
11 captured options turns. Investigation dated July 25, 2026. A live production before-and-after was not recorded. This result is not a savings forecast for your system.
Read the method and its limitsWhere we look
We inspect the assembled requests and actual tool responses, then check the changes against the work the agent needs to do.
- 01Prompts & context
- Send the instructions and context the task needs. Remove duplicated or irrelevant material without removing the rules that make the answer useful.
- 02Tool output
- Inspect what tools send back to the model. Trim repeated fields, oversized responses and data the next step never uses.
- 03Caching
- Check what is being reused and what is being billed again. Measure cache reads, writes and misses across the full turn.
- 04Models & calls
- Test which steps can use a cheaper model, and which calls can be avoided. Compare the result, retries and latency before keeping a change.
How the work runs
We measure the baseline, implement the changes, and check the result on the work your agent actually does.
- 01 / Measure
See what the task costs.
Benchmark representative turns and trace expensive steps. Keep model spend, failed attempts and human review visible.
The output: a baseline and priorities.
- 02 / Change
Put the fixes to the test.
Implement the agreed changes and compare them on the same tasks. Keep cost and output quality in the same review.
The output: changes with a replay comparison.
- 03 / Check
Measure on live work.
After rollout, check the result on live traffic. Watch for costs moving into retries, delays or extra review.
The output: a view of what changed in use.
Practical guides
Understand your agent’s costs and what to check first.
Four checks for wasted tokens
System prompt, tool output, caching and model choice.
Operating costWhat an accepted result really costs
Include the retries, review and work around the model.
Scoping an engagementWhat changes the project budget
Understand scope, integration and ongoing operation.
Before we talk
Do we need to change model providers?
That depends on the workload. We first look at the prompts, tool responses, caching and calls in the system you already run. A model change is one option to test.
How do you check that cheaper results are still useful?
We agree on representative tasks and what counts as an acceptable result, then compare changes against that baseline. Cost, completion, latency and review effort need to be read together.
What should we bring to the first call?
An overview of the workflow, which models it uses, and where the costs are causing concern. We can discuss the logs or traces needed to scope the work; you do not need to send confidential data to book.
What does the work cost?
The scope depends on your workflow, the evidence available and the changes involved. We agree on the work and fee before starting.
Olypsis / Let’s look at the work
Know where your
AI budget goes.
Tell us what the agent does and where the costs are causing concern. We’ll discuss what it would take to investigate.