How Claude Code, Codex, and OpenCode Can Cut Token Usage by Up to 80%

Token billing follows the formula

Total Cost = Input Tokens × Input Price + Output Tokens × Output Price

, with input tokens accounting for 70‑90% of the expense, making input optimization the primary cost‑saving lever.

1. Claude Code (Claude) Optimizations

File filtering : create a .claudeignore file (syntax like .gitignore) at the project root to exclude build artifacts, logs, caches, and other irrelevant files. Example content:

# Dependency and build folders
node_modules/
.dist/
.build/
__pycache/
# Lock files and logs
*.lock
package-lock.json
*.log
# VCS and IDE files
.git/
.idea/
.vscode/
# Assets
*.png
*.jpg
*.svg
*.ico
.cache/
coverage/

Result: a single interaction drops from 150 k to 60 k tokens (≈60% reduction).

Context compression : use /compact manually or enable auto‑compact via /config (e.g., Auto‑compact enabled). This trims long dialogues, keeping only essential code changes. Result: 25 k → 3 k tokens (≈88% saved).

Documentation drive : add a CLAUDE.md at the root describing the stack, directory layout, and common commands, so the model can retrieve project context without scanning files.

Memory management : store recurring information with /memory (e.g.,

/memory 项目用 Next.js 14 + TypeScript，接口规范见 docs/api.md

) and retrieve it via /memory list or delete with /memory delete [key]. Saves >40% of repeated input.

Plan mode : press Shift+Tab to let the model propose an execution plan before running code, avoiding wasted tokens from trial‑and‑error. Saves >20% of invalid output.

Output trimming : enable concise tool output with /config to strip ANSI colors, progress bars, and empty lines; truncate long logs to error stacks only. Example: npm test output reduced from 25 k to 2.5 k tokens (≈90% saved).

Model switching : select appropriate model per task ( /model haiku for simple syntax, /model sonnet for complex architecture, /model opus only when necessary). Cost reduction ranges from 30% to 80%.

2. Codex (GitHub Copilot) Optimizations

Limit the IDE’s maximum file context (set GitHub Copilot → Max File Context to 3–5 files) to prevent full‑project scans, cutting input tokens by >50%.

Use short inline prompts instead of verbose natural‑language requests, e.g., replace a long description with // Node.js Express 登录接口 JWT bcrypt, reducing input tokens by >40%.

Disable unnecessary features such as real‑time suggestions and automatic multi‑file indexing, and adopt a per‑file development style to shrink context size by >60%.

3. OpenCode (Self‑hosted) Optimizations

Configure .opencodeignore similarly to .claudeignore to filter out irrelevant files.

Set precise context limits in config.json (e.g., "input_limit": 128000, "output_limit": 80000) to fully utilize model capacity and avoid truncation, saving >30%.

Manually clear history with /clear and start new sessions for separate tasks to prevent context bloat, yielding >50% reduction in wasted tokens.

4. Comparative Savings Overview

Across the three platforms, token‑saving techniques achieve the following approximate reductions:

File filtering: 60%–80%

Context compression: 50%–88%

Documentation drive: 30%–50%

Memory caching: 40%–60%

Plan mode: 20%–40%

Output trimming: 70%–90%

Model switching: 30%–80%

Context limit tuning: 30%+

5. Practical 10‑Step Token‑Saving Checklist

Create .claudeignore / .opencodeignore at the project root using the provided template.

Add a CLAUDE.md (or README_OPENCODE.md) describing the tech stack, directory structure, and commands.

Enable auto‑compact with /config → Auto‑compact.

For long conversations, manually invoke /compact and periodically clear history.

Store project configuration and standards in /memory to avoid repeated input.

Use Plan Mode ( Shift+Tab) for complex tasks before execution.

Switch models per task: simple → Haiku/low‑cost, complex → Sonnet.

Turn off unnecessary auto‑features (real‑time suggestions, full‑project scans).

Develop in separate sessions or files to keep context size small.

Regularly check token usage with /usage to locate remaining “black holes”.

6. Key Reminders

Prioritize optimizing input: exclude as many files as possible.

Prefer over‑exclusion; excluded files can be manually pasted if needed.

Clean up long dialogues and multi‑task histories promptly.

Match the model tier to the task complexity; avoid defaulting to the most expensive model.