Cut your LLM bill by 40-70% without cutting quality.
Vendor-neutral LLM cost audit and re-architecture for production AI apps. Prompt caching, model routing, semantic caching, token compression, batch inference. £8K, 5-day audit — average client saves 47% of their LLM spend inside 30 days.
What is LLM cost optimisation?
LLM cost optimisation is the practice of reducing the per-request cost of production AI apps without degrading the user experience. In 2026, the four biggest cost levers are: prompt caching (Anthropic + OpenAI both support first-class prompt caching now — 90% cheaper reads), model routing (route 60-80% of traffic to smaller / cheaper models like Haiku or GPT-4.1 Mini and reserve the frontier model for the 20% of hard cases), semantic caching (dedupe recurring queries at the app layer with a vector-similarity threshold), and prompt compression (LLMLingua-style token reduction on the input side). Empyreal Infotech runs a fixed £8K, 5-day audit that measures your current spend by route, models the four levers against your traffic, ships a re-architecture plan, and implements the top three wins in a follow-on sprint. Average 2026 client cut their LLM bill by 47% inside 30 days.
What you get, every engagement.
Token audit by route
Every audit starts with a real measurement: token spend per API route, per model, per tenant, per time window. Not a vendor bill; a per-request accounting so you know which routes are the actual money.
Prompt caching + model routing
Anthropic prompt caching (90% cheaper reads), OpenAI prompt caching (50% cheaper reads), model routing (Haiku / GPT-4.1 Mini / Gemini Flash for the easy 70%, Sonnet / GPT-5 / Gemini Pro for the hard 30%). Fallback chains with cost caps enforced.
Semantic + response caching
Vector-similarity caching at the app layer (Redis + PGVector). Recurring queries get answered from cache, cache-hit rate reported per route. Typical B2B SaaS see 25-40% cache-hit rate on customer-facing chat surfaces.
Batch + async patterns
Anthropic + OpenAI Batch APIs (50% cheaper, 24h SLA) for non-interactive workloads. Background job re-architecture so nightly digests, embeddings, and moderation runs use the batch tier instead of realtime.
The LLM cost optimisation engagement, week by week.
- 01
Wire a temporary cost-tracking middleware, measure 5-7 days of real traffic, model each cost lever, produce a 30-page audit with cost-vs-quality trade-offs per route and a ranked implementation plan.
Audit week (£8K). Wire a temporary cost-tracking middleware, measure 5-7 days of real traffic, model each cost lever, produce a 30-page audit with cost-vs-quality trade-offs per route and a ranked implementation plan.
- 02
Ship the three wins that need no product change: prompt caching, model routing on obvious routes, batch API on background jobs. Typical cost cut inside week 2: 25-40%.
Quick wins. Ship the three wins that need no product change: prompt caching, model routing on obvious routes, batch API on background jobs. Typical cost cut inside week 2: 25-40%.
- 03
Ship the app-layer semantic cache with the similarity threshold you'll actually accept. Add LLMLingua-style prompt compression on the largest inputs. Typical cumulative cut by end week 4: 40-60%.
Semantic caching + prompt compression. Ship the app-layer semantic cache with the similarity threshold you'll actually accept. Add LLMLingua-style prompt compression on the largest inputs. Typical cumulative cut by end week 4: 40-60%.
- 04
Cost dashboard wired to Datadog / Grafana. Per-tenant cost budgets with alerts. Regression gate in CI so a bad prompt change can't blow the budget silently.
Measurement + guardrails. Cost dashboard wired to Datadog / Grafana. Per-tenant cost budgets with alerts. Regression gate in CI so a bad prompt change can't blow the budget silently.
- 05
£3K/month for monthly cost reviews, model-drift response when providers ship new pricing, and a quarterly re-audit.
Optional retainer. £3K/month for monthly cost reviews, model-drift response when providers ship new pricing, and a quarterly re-audit.
Questions we get about LLM cost optimisation, with real answers.
The average 2026 client cut their LLM bill by 47% inside 30 days. The range across 9 audits was 28% (already-optimised app) to 71% (early-stage app with no caching + one big model on every route). The £8K audit produces a numbers-backed estimate for your specific traffic pattern before you commit to the build.
All of them, vendor-neutral. Anthropic (Claude Sonnet, Haiku, Opus), OpenAI (GPT-5, GPT-4.1, Mini, o-series), Google (Gemini Pro, Flash, Ultra), Groq, DeepSeek, together.ai. If you're on a single provider today, the audit will also cover whether multi-provider routing is worth the operational complexity for your traffic.
The audit measures quality alongside cost using either your existing evals or a temporary eval harness we build. Any route where the cheaper model degrades quality below your bar stays on the frontier model. The 47% average savings is post-quality-check.
Yes. Both providers support prompt caching in 2026, with different pricing and cache semantics. Anthropic: 5-min ephemeral cache, 90% cheaper reads, 25% more expensive writes. OpenAI: automatic caching on 1024+ token prompts, 50% cheaper reads. Cache design (which prefixes to cache, how to structure prompts to maximise hits) is a core audit deliverable.
Yes — LLMLingua-2 and similar techniques where the input is long and repetitive. Typical wins are 30-50% token reduction on RAG-heavy prompts. We don't compress user turns; we compress system prompts, retrieved context, and tool descriptions.
Yes. Redis + PGVector (or Weaviate / Chroma if you already run one) at the app layer, with a configurable similarity threshold. Typical B2B SaaS see 25-40% cache-hit rate on customer-facing chat. Cache invalidation strategy designed per route.
Yes — Anthropic and OpenAI both offer 50% cheaper batch APIs with a 24h SLA. Any workload that doesn't need realtime (nightly digests, moderation, embedding generation, RAG re-indexing) can move to batch. Usually the fastest single win in the audit.
£8K fixed. You walk away with the audit deck on Friday whether you build with us or not. Roughly 30% of 2026 audit clients took the deck and implemented it with their own team. We're still available for PR review at £250/hr if you want a senior eye on the work.
Send a 5-line brief. That's it.
5 lines: which providers you're on, your rough monthly LLM spend, your traffic pattern (interactive vs background vs mixed), and your biggest cost concern. Mohit replies inside 24 hours with a savings estimate + audit slot.