7 Practical Ways to Reduce Claude Code Token Usage#

Original | Raw

Author: Kanwal Mehreen (KDNuggets) Published: 2026-05-03

Summary#

Practical guide to reducing claude-code token costs. The core insight: you’re not paying for the prompt you typed — you’re paying for the entire context window, which includes prior messages, files read, tool outputs, and persistent files like CLAUDE.md. Most token waste comes from messy context, not bad prompting.

Key Takeaways#

1. Switch Models by Task Complexity#

  • Opus (5× Sonnet cost): complex multi-file architecture, gnarly debugging
  • Sonnet: day-to-day — tests, edits, explaining, refactoring
  • Haiku: mechanical — lookups, formatting, renaming, repetitive tasks
  • Use /effort to reduce thinking budget on straightforward tasks

2. Keep CLAUDE.md Small and Useful#

  • CLAUDE.md loads on every single turn — a 5,000-token file costs 5,000 tokens per message
  • Put stable instructions: test commands, package manager, formatting rules, architectural constraints, directories to avoid
  • Keep it lean — no meeting notes, design history, or long implementation guides
  • Works best as a lookup table, not a brain dump

3. Delegate Verbose Work to Subagents#

  • Subagents run in isolated context windows — verbose output stays contained, only summary returns
  • Not automatically cheaper — startup overhead (prompts, tool definitions, round trips) makes them wasteful for small tasks
  • Rule: use subagents when saved main-context clutter > startup overhead

4. Point Claude to Exact Files and Line Ranges#

  • Vague tasks trigger expensive exploration (opening files, dead ends, reconstructing context)
  • Specific: "Compare src/auth/session.ts lines 30-90 with src/api/login.ts lines 10-60"
  • Use plan mode (Shift+Tab) before expensive operations — review plan, cut unnecessary steps, then execute

5. Use /compact Proactively#

  • Compact before context gets overloaded, not after Claude starts forgetting
  • Best timing: after exploration is done and important parts are clear, before next phase
  • Early compaction produces better summaries than late compaction

6. Check /context Before Optimizing#

  • /context shows what’s actually consuming the window
  • Biggest improvements often come from spotting one “quiet offender” riding along every turn
  • Inspect first, then optimize — don’t optimize blindly

7. Keep Tooling Setup Simple#

  • More connected tools = more context overhead
  • Only load integrations that solve a real repeated problem

Key Quote#

“Stop thinking only about prompts and start thinking about context architecture.”

Connections#

See Also#

Backward Claude Code Docs Continuum Memory Architectures Forward