本文由 AI 分析生成
建立時間： 2026-05-25 來源： https://x.com/DeRonin_/status/2054235707791778034
Summary
A detailed breakdown of how one developer cut their monthly AI coding bill from 312 by applying context discipline, multi-model routing, and prompt caching — without reducing shipping velocity. The core insight is that most token waste comes from re-sending unchanged context on every turn, not from model pricing itself. The guide proposes routing most serious coding work to Kimi 2.6 (a cheaper alternative matching Sonnet quality) and reserving premium models for the 10% of architectural decisions that truly compound.
透過上下文管理、多模型路由與提示詞快取，將 AI 編碼月費從 312，降幅約 93%。核心洞察是：大部分 token 浪費來自每輪都重送未變動的上下文，而非模型本身定價問題。作者建議將日常編碼主力切換至 Kimi 2.6（品質接近 Sonnet 但便宜 6 倍），Opus/GPT-5 僅用於架構決策。
Key Points
- Context discipline is more impactful than model selection: sending 80k tokens of “just in case” files per turn is the primary driver of bloated bills
- Prompt caching reduces cached-prefix token costs to ~10% of normal input cost; available in Claude Code, Cursor, and Aider
- Multi-model router: Opus/GPT-5 for architecture (10% of tasks), Kimi 2.6 for implementation/debugging/refactoring (bulk), Haiku for lint/format, local models (Qwen3 via Ollama) for boilerplate
- Benchmarks show Kimi 2.6 matches Sonnet 4.6 quality at 3-4x lower cost across refactoring, CRUD, debugging tasks
- Summarize context aggressively: after 10-15 turns, compress 200k-token sessions to 5k-token summaries before starting a new batch
Insights
The framing “you’re not paying for tokens, you’re paying for context” reframes the optimization target from model pricing to information architecture. The “graduated skill pattern” (saving discovered workflows as SKILL.md files to skip discovery on subsequent runs) parallels the bot_vault skill system exactly. The 30x cost multiplier between message 1 and message 30 in a long conversation is rarely internalized by developers until they measure it.
Connections
Raw Excerpt
Context discipline is the lever. Model selection is downstream of it. Once you internalise this, every section below makes sense.