How effective is Claude Code's prompt caching? Shows cache hit ratio and illustrative API cost savings across your session history.

cc-cache — Claude Code cache efficiency

  Sessions:       755
  Total tokens:   23.2B
  Input (fresh):     8.7M tokens (0.0%)
  Cache write:     783.7M tokens (3.4%)
  Cache read:       22.4B tokens (96.5%)
  Output:           20.1M tokens (0.1%)

  Cache hit ratio:  96.6%
  ███████████████████████░

  At API rates (illustrative):
    Without cache:   $293.1K
    With cache:       $41.6K
    Saved:           $251.5K  (85.8% saved)

  ⓘ  Claude Max plan is flat-rate. Costs above are illustrative API equivalents.

npx cc-cache              # Cache efficiency and illustrative savings
npx cc-cache --all        # Include subagent session files
npx cc-cache --json       # JSON output

• Cache hit ratio — percentage of input tokens served from cache vs. freshly processed
• Token breakdown — input / cache write / cache read / output split
• Illustrative savings — what prompt caching would save at Claude API rates
• Monthly efficiency — month-by-month cache performance
• By project — which projects benefit most from caching

Claude Code on Max plan is flat-rate — you don't pay per token. The cost figures use Claude API Opus pricing as an illustrative reference to show the value of prompt caching.

The methodology:

• Without cache: all input tokens billed at full input rate
• With cache: cache_read tokens at 10% of input rate, cache_write at 125% of input rate

Reads session file content to extract token usage metadata. No content is transmitted. Everything runs locally.

Drop your ~/.claude folder into cc-cache on the web — no install required. Note: large histories (5+ GB) may take 1–2 minutes to process.

Part of cc-toolkit — 104 free tools for Claude Code