Tokenminning
Tokenminning is the deliberate practice of reducing large language model (LLM) token consumption while preserving useful output quality. Teams treat inference tokens as a finite resource and optimize where spend does not convert to measurable value.
Press coverage, job posts, and internal docs may say token minimizing — searches for that phrase describe the same practice. This wiki standardizes tokenminning as the engineering term for instrumented, attributed cost control: meter usage, attribute spend per feature, and enforce inference budgets in production.
It is the counter-move to tokenmaxxing: maximizing raw AI usage, often tracked on leaderboards by token volume rather than shipped outcomes. When inference bills caught up with that habit, tokenminning became standard engineering practice — meter first, then trim prompts, route models, and cap agent loops.
Where should you start?
Follow the branches — every endpoint links to a guide on this wiki.
Each branch optimizes for a different cost model: remove inference entirely · stretch subscription quota · trim per-token API spend.
View paths as plain text
- Remove AI entirely: Can you stop using AI? → Yes → Remove LLMs from your stack
- Subscription / IDE: Can you stop using AI? → No → How do you pay for tokens? → Subscription → Prompt hygiene → IDE guides → Where to start
- API billing: Can you stop using AI? → No → How do you pay for tokens? → API per-token → Metering first → Model routing → Prompt hygiene → Prompt caching → Context compaction → Output and RAG → Semantic caching → Local inference → Self-host a model
What’s on this wiki
| Section | Teaser |
|---|---|
| Practice | Implementation guides ordered by leverage — metering, prompt hygiene, model routing, caching, and context compression. Each guide includes expected savings ranges and anti-patterns. Start with Where to start. |
| Constitution | Engineering law for production AI stacks — immutable metering, model routing rules, session caps, prompt lint rules, and CI blocks. Load-bearing guardrails, not style guides. |
| Manifesto | The philosophical case for treating inference as a scarce resource — why tokenmaxxing fails, what tokenminning advocates, and how unit economics should shape architecture. |
| Concepts | Definitions for token economics — input vs output tokens, context inflation, model selection, and the full tokenminning definition. |
| Agents | Eve agent execution costs on Vercel Sandbox — diagnose idle spend, estimate sandbox vs self-host execution, and explore supported pilot interest. |
| Self-hosting | When on-prem inference beats cloud APIs — GPU metering, workload fit, iwant recipes, and deep dives on Ollama, vLLM, SGLang, llama.cpp, TGI, and LocalAI. |
| IDEs | Per-editor guides for Cursor, Copilot, Claude Code, Cline, and others — same optimization sequence, different controls for subscription vs API billing. |
Docs are also available via MCP and llms.txt for agents and IDE tooling.
Saas
Platform charges can grow independently of model usage. Saas covers how to reduce your Snowflake bill, reduce your Salesforce bill, and evaluate self-hosting Dagster, including the operating costs of open-source alternatives.