LLM Cost Optimisation
You set the cost per answer
Most AI bills grow with context length. So you choose the length, and the bill follows.
- Model routed by task class
- Context kept lean, one task per session
- A cap on each session, a weekly baseline for trends
- Stable prompt openings, served from the cache
Example record
Where your tokens went this week
Spend per task class, against last week.
A looping session stops at its cap, so the baseline only shows trends.
- Normal
Classification on a small model
within the baseline
- Normal
Architectural questions via the index
answered from the index
- Stopped
A session looping on one test
stopped at its spend cap
- Refused
A repository sweep
25,000 to 50,000 tokens
Where the money goes
Everything in the context window is re-sent on each turn. Cost grows with session length.
The largest model used for classification work
Whole repositories loaded to answer one question
Sessions that run long and loop
Spend nobody reads until the invoice
Capabilities
What moves your bill
Measured on our own delivery work, then applied to yours.
Model
What we fix, and what you decide
What we fix
- F1Spend measured per task class
- F2A turn and spend cap on each session
- F3A weekly baseline
- F4One task per session
What you decide
- C1Models and providers
- C2The budget and its alert threshold
- C3Which tasks get the large model
Next step
Send us last month’s invoice.
We break it down by task class. You take a costed breakdown to your next budget review.