cheaper. same output.
3 changes that cut your Claude Code bill. Caching, batching, model routing, prompt compression.
Each strategy stacks. Apply all three and you hit 80% savings.
Shorthand reasoning format that compresses chain-of-thought into dense, token-efficient notation. Same quality output, fraction of the cost. The model understands it perfectly.
Every MCP server loads its full schema into context on every call. A CLI tool loads nothing — just runs and returns. Replace MCP tools with lightweight CLI wrappers wherever possible.
Most sessions carry dead context — old file contents, stale tool schemas, irrelevant history. Aggressively prune what the model sees. Only load what it needs for the current task.
Same task. Optimized vs unoptimized session.
*Estimates based on community benchmarks and documented results.
| Method | Token Saving | Impact | Effort |
|---|---|---|---|
| Caveman Tokenization | –75% | High | Low |
| CLI over MCP | –97% | High | Medium |
| Context Pruning | –98% | High | Low |
| Cache Token Drain Fix | –90%* | Medium | Low |
| Built-in API Credits | $20–$100/mo | Free money | Zero |
No setup required.
Fixes a known issue causing 10–20x token drain from broken cache invalidation.
github.com/Rangizingo/cc-cache-fixPro plan includes $20/mo. Max plan includes $100/mo (5x). Check Settings > Usage.
Most people don't know these exist.Drop it in and start saving on the next session.
# CLAUDE.md — Token optimization output-style: default # Caveman tokenization (75% fewer reasoning tokens) Use shorthand reasoning: abbrev words, drop filler, dense notation. Think: step→, result=, b/c=, w/=, cfg=, init=, upd= # CLI over MCP Prefer CLI tools over MCP schemas wherever possible. MCP: load full schema per call. CLI: zero schema overhead. # Context pruning Only load files needed for current task. Prune stale tool output, old file contents, irrelevant history.