Cost Optimization & Token Management

Claude Code charges based on the tokens you use. Here's how to identify where tokens are leaking and reduce them.

---

Checking Current Token Usage

/context

Displays token usage broken down by category in a bar chart format.

Category	Meaning
System Prompt	System instructions such as CLAUDE.md
System Tools	Connected MCP servers, tool definitions
User/Assistant	Actual conversation content

---

TOP 5 Token-Heavy Patterns

1. Always Keeping Heavy MCPs On

MCPs that handle large volumes of data, like Notion or Linear, consume 10% or more of System Tools tokens just by being active.

Fix: Enable only when needed
/mcp → deactivate/activate the relevant server

2. CLAUDE.md Becomes Too Long

Putting everything in the root CLAUDE.md wastes thousands of tokens every turn.

Fix: Split into per-folder CLAUDE.md files
Root: Only core rules, 50–100 lines
Subfolders: Only domain-specific rules → loaded only when working in that folder

→ See CLAUDE.md Hierarchy for details

3. Developing Multiple Features in One Session

As context accumulates, the history Claude must process grows, making it progressively slower and more expensive.

Fix: One session = one feature principle
Always run /clear after completing a feature

4. Analyzing Large Files Directly in the Chat

Pasting thousands of lines of JSON or log files into the chat consumes tens of thousands of tokens at once.

Fix: Script offloading
Ask Claude to write an analysis script,
run it locally → pass only the summarized results back to Claude

→ See Script Offloading for details

5. Overusing Subagents

Each subagent uses its own independent context. Running 3 in parallel can consume up to 3× the tokens.

Fix: Only run in parallel when tasks are truly independent
If not independent, a single session is far more efficient

---

Cost Strategy by Model

Claude Code lets you switch models mid-task with /model.

Model	Relative Cost	Suitable Tasks
Opus	Highest	Architecture design, complex bugs, critical decisions
Sonnet	Medium	General coding, parallel subagents
Haiku	Lowest	File exploration, simple renaming, quick questions

Pinning the subagent model:

# Main session: Opus (judgment), subagents: Sonnet (execution)
export CLAUDE_CODE_SUBAGENT_MODEL="claude-sonnet-4-5"

---

3-Step Context Management

Situation	Command	Effect
Context is over 70% full	/compact	Compresses down to a core summary
A single feature is complete	/clear	Full reset, starts a new session
Need to carry context across sessions	Update TODO.md then /clear	Context lives in files, memory is cleared

---

Cost Monitoring Checklist

Before starting work:

• [ ] Check current token usage with /context
• [ ] Disable unused MCP servers
• [ ] Select the right model for the scope of work

During work:

• [ ] /clear after each feature is complete
• [ ] Offload large files via scripts
• [ ] Use think hard only when truly necessary (increases internal reasoning tokens)

After work:

• [ ] Move recurring patterns from root CLAUDE.md into a folder-level CLAUDE.md
• [ ] Automate frequently used tasks with custom commands or hooks