Summary

Claude counts tokens, not messages — and the quadratic growth of token costs in long conversations (Total = S × N(N+1)/2) means message 30 costs 31x more than message 1. This guide provides 10 concrete habits to reduce token consumption: editing prompts instead of follow-ups, starting fresh every 15-20 messages, batching questions, using Projects for recurring files, leveraging memory settings, disabling unused features, using Haiku for simple tasks, distributing work across the day, working during off-peak hours, and enabling overage as a safety net.

Claude 計算 token 而非訊息數,長對話的 token 成本呈二次方增長(第 30 條訊息的成本是第 1 條的 31 倍)。本文提供 10 個具體習慣降低 token 消耗,包括編輯提示詞而非追加、每 15-20 條訊息開新對話、批次提問、善用 Projects 快取等。

Key Points

  • Token cost formula: Total = S × N(N+1)/2; at 500 tokens/exchange, 30 messages = 232k tokens vs 500 for message 1
  • Fix: click Edit on original message to replace instead of appending to conversation history
  • Start fresh every 15-20 messages; summarize context → copy → new chat → paste as first message
  • Batch multiple questions into one prompt: saves context reloads and gives Claude better full-picture answers
  • Projects feature caches uploaded files; recurring PDFs/docs uploaded once don’t re-tokenize
  • Rolling 5-hour window (not daily reset); peak hours (5-11am PT weekdays) burn limits faster as of March 2026
  • Haiku mental model: Haiku for drafts/simple tasks, Sonnet for real work, Opus for deep thinking

Insights

The 98.5% statistic (one developer found 98.5% of tokens went to re-reading history, only 1.5% to actual output) is extreme but directionally accurate for very long sessions — it quantifies why context discipline matters more than model selection. The peak-hours multiplier (same query costs more during peak hours) is underappreciated, especially for non-US users who may be hitting US peak hours during their own afternoon. Note: this file is a duplicate of the non-”1” version with identical content.

Connections

Raw Excerpt

Claude doesn’t count messages. It counts tokens.