LinkedIn publication text
Cutting LLM Cost by 70% Without Losing Quality
🚀 Can you cut LLM cost by 70% without losing quality?
R. Bhardwaj published a six-lever proposal to do exactly that. I checked each part against primary sources and turned the analysis into a 41-slide infographic, with 13 original figures and 55 audited references.
What I found:
- 🔵 The thesis holds: routers and cascades matched the strongest model at a fraction of the cost (FrugalGPT: up to 98%; RouteLLM: over 2×).
- 🟢 Savings multiply, they don't add up: 0.85 × 0.60 × 0.75 × 0.85 × 0.90 = −70.7%. Routing alone accounts for 34 of those points.
- 🟣 Prompt caching and batch are the lowest-risk levers: cached reads at 0.1× the input price and a 50% batch discount at the major providers.
- 🟠 The 70% figure is illustrative: once you add routers, judges, embeddings and evaluation, a plausible net saving is 61–68%.
- 🔴 "Same quality" needs proof: on a 500-item eval the noise is ±2.4 points, larger than the reported gains.
- 🟡 Semantic caching is the riskiest lever (false hits, stale answers, cross-tenant leaks) and should come last.
The metric I recommend: cost per successful task. A model that is 10× cheaper but succeeds half as often is only 5× cheaper per success.
Suggested roadmap: measure first, take the low-risk wins (dedup, loop caps, prompt caching, batch), then add routing and caching behind quality gates.
📎 The full infographic is in the PDF. Which of these levers do you already run in production?
© Aridio Silva @aridiosilva · October 2026