LLM cost optimisation

When this helps

This work suits teams whose LLM bills are rising, whose workflows make repeated calls, or whose prompts carry more context than each task needs. It is also useful when a smaller model or another route may fit some requests, but quality varies across mixed cases and needs checking before a change.

Measure cost per successful task

Start with representative requests and define what counts as a successful result for each workflow. Review token use, model calls, retries and failures together. A lower token count is not useful if more requests need another attempt or produce unusable output.

Separate request types where context, output length or quality requirements differ. This shows which parts of the workload drive spend and avoids treating an average as if every task behaved the same way.

Find avoidable context and repeat work

Trace how prompts are assembled and where the same information is sent again. Check whether context can be narrowed, shared work reused, or caching applied without serving stale or incomplete information.

Review retries and chained calls to understand why they happen. Some are needed for quality or reliability; others repeat work due to routing, prompt structure or unclear stopping conditions. Changes must preserve the checks that protect the workflow.

Match model routes to request needs

Compare current model API costs with suitable self-hosted options against the actual task mix. A route can use different models or optimised endpoints for different request types when the quality bar and failure handling are clear, keeping more demanding cases on a suitable path.

Test proposed routes on the same representative inputs. Assess output quality alongside cost, latency and errors. Include routine and difficult cases so the comparison shows where each route works and where a fallback is needed.

Put verified changes into the workflow

Select experiments based on the baseline and the risks of changing the workflow. Compare results under consistent request conditions so lower spend can be considered alongside successful completion and quality.

Implement changes that pass the agreed checks, then hand over the relevant configuration and monitoring approach. Scope, price and timeline are agreed for the workload before implementation begins.

What you receive

• Baseline of model calls, token use, cost and successful task outcomes for representative requests.
• Review of context size, repeated calls, retries, caching opportunities and routing choices.
• Comparison of selected changes using consistent inputs and quality and error checks.
• Implementation of agreed changes with configuration, monitoring guidance and a rollback plan.

Common questions