Author(s): Ashish Nishad

Originally published on Towards AI.

There are a hundred “10 ways to save on tokens” articles out there. This is what I actually built, on a real internal tool, and the one change that did most of the work.

I want to skip the generic advice for a second. You’ve probably already read “shorten your prompts” and “cache your context” a dozen times. Some of it’s useful. Most of it treats token cost like a checklist instead of what it actually is: an architecture decision.

After the introduction, the article explains that the main mistake most teams make is routing every request through the same most-capable model, which quietly becomes an architecture-driven cost decision. The author’s fix is to treat model selection like per-request routing: classify each request by how much reasoning it truly needs, then send routine tasks to a cheaper model and genuinely hard tasks to a more capable one. On a real internal automation tool, this routed model choice reduced costs by roughly 20–40%, with week-to-week variability reflecting the changing mix of request types—useful visibility rather than an average number. The author also notes supporting techniques with additional leverage (prompt caching, trimming conversation history, and using retrieval of relevant chunks instead of dumping full documents), and closes with practical guidance for engineers to start by measuring traffic and cost, classifying their own workload, routing deliberately with logging, and expecting savings to vary as information about the system.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI