Loading…
service
You prototype with GPT-5, it works well, and then at 50,000 queries a month your bill hits $8,000 and you can't figure out where it's going. No single endpoint looks obviously wrong. But the total is unsustainable. This is the standard pattern for AI products that scaled faster than their infrastructure was optimised.
We audit your token usage, identify the 3–5 highest-cost patterns, and implement targeted fixes. We measure quality before and after each change with an eval suite. Average reduction from a cost audit engagement: 40–60% spend reduction. The audit alone takes two weeks. Implementation of the top fixes takes another two to four.
Tell us what you're building.
Not every lever applies to every product. The audit determines which ones move the needle for your specific usage pattern.
Route simple queries (classification, extraction, summarisation) to GPT-5 mini or Claude Haiku 4.5 at roughly 1/10th the cost of the flagship model. Reserve GPT-5 or Claude Sonnet for tasks that actually require complex reasoning. This is a single configuration change with no quality impact on the simple queries. Typical result: 30–50% cost reduction before touching anything else.
Cache LLM responses by embedding similarity rather than exact match. A question that is 95% similar to a cached one returns the cached answer immediately, no API call, no latency. For products where users ask similar questions repeatedly (support bots, knowledge assistants, internal tools), this cuts 20–40% of LLM calls. Cache entries expire on a TTL you control.
Strip redundant context from system prompts, compress long-form instructions with LLMLingua or a similar approach, and move static context that never changes into a vector lookup rather than pasting it inline. Done carefully (with quality measurement before and after), this reduces input token count by 20–50% with no degradation in output quality.
Batch non-real-time requests (document processing, report generation, overnight analysis jobs) and submit them through the OpenAI Batch API or equivalent. This cuts per-token pricing by 50% on eligible workloads. If your product has any background processing, this is often the fastest win after model routing.
Audit your RAG retrieval to find where you are stuffing 8,000 tokens of context when 2,000 is enough. Oversized context windows inflate costs, slow responses, and dilute the relevant information. Tuning the retrieval window is frequently the single largest cost win in a RAG system, and it also improves answer quality.
The audit is a two-week engagement that gives you a clear picture of where your money is going and exactly what to fix first.
30-day usage log analysis from your OpenAI and Anthropic dashboard exports
Token breakdown by endpoint, model, and prompt template
Top-10 most expensive query patterns, ranked by monthly spend
Written report with prioritised fixes ranked by estimated impact vs. implementation effort
Optional: we implement the top three fixes as part of the engagement
Cutting a bill is easy if you do not care about quality or whether the number holds next quarter. The discipline is in proving the savings, protecting the product, and keeping spend from creeping back.
Baseline before we touch anything. We instrument a month of real usage and convert it into cost per request, per feature, and per prompt section. That signed baseline is what every later claim is measured against, because a savings number nobody can reproduce is just a story.
Every change is gated on quality. Before a cheaper model or a compressed prompt ships, it has to clear an evaluation set built from your real traffic. If a change drops quality below the threshold you approved, it does not ship. Cost reduction that quietly degrades the product is not a win, and we will not hand you one.
Attribution your finance team can trust. Each fix is reported in the same units the baseline used, with the method documented, so the before-and-after reconciles to the provider invoice. The deliverable is a number your controller can re-derive next quarter without us in the room, which is what turns a one-time saving into standing credibility.
Instrumentation so it does not creep back. Spend drifts up again when a new feature ships without cost awareness or a prompt quietly doubles in length. We leave behind per-feature dashboards and spend alerts, so the next regression is a Tuesday email rather than a quarterly surprise, and your team can hold the line without another engagement.
A cost optimisation engagement pays back when the spend is high enough to justify it and when the system is stable enough to optimise.
Teams spending under $500 per month on LLMs
At that spend level, the audit cost will not pay back within a reasonable timeframe. Come back when you hit $2,000–$3,000 per month and the optimisation savings become meaningful.
Projects where latency is the only constraint
Some optimisations trade latency for cost, semantic caching adds a lookup step, model routing may use a slower model path, batching introduces delay. If every request must complete in under 500ms and cost is not a concern, the tradeoffs don't work in your favour.
Anyone who needs a full system rebuild
Cost optimisation works on an existing system. If your architecture is fundamentally broken, wrong model choices, bad retrieval design, no batching at all, the right answer is a rebuild, not incremental tuning of a poor foundation.