You’ve heard of tokenmaxxing? We are now in the age of tokenminning, where enterprises are getting spooky bills from their cloud LLM providers, larger than a human developer’s salary.
Here are five easy ways to dramatically reduce token cost without sacrificing output.
1. Model routing
Not every query deserves the flagship model. Classification, extraction and formatting go to a cheap model. Architecture and genuine reasoning go to a frontier model. Execution on that architecture and reasoning goes to the mid tier. Most teams pay frontier prices for work a mini model handles identically. This is usually the single largest saving available, bigger than any context trick on this list.
2. Cache the static information
System prompt, tool schemas, coding standards and architecture docs, ordered before the dynamic content and cached. Get the ordering wrong and you pay full price on every call for information that never changed. Get it right and it’s free money.
3. Run deterministic commands yourself
Don’t ask the model how many files changed. Run git diff --stat and hand it the
output. Every deterministic computation offloaded to a tool is a reasoning token
saved for the part of the task that actually needs reasoning.
4. Compress before it enters context
Raw logs, stack traces and full diffs shouldn’t be piped through unfiltered. A 20 line semantic summary of a 4,000 line log carries the same signal at a fraction of the cost. This is the idea behind dedicated compressors. I use headroom, and it is great.
5. Narrow, staged retrieval
Don’t pull 30 documents when only 3 are necessary, and don’t re-search the repo on every query. Escalate to retrieval only when confidence actually drops, not automatically.
Each of these maximises the value of every token sent. Valuemaxxing is tokenminning.
The fixed AI budget you argued for can now get 50% more output for the same cost, before hitting limits. Build these five into your harness and the savings at scale are considerable.