PromptCrunch vs provider prompt caching

These are different tools for different traffic. Sometimes caching is all you need, and we'll say so. In fact, if your request uses caching, PromptCrunch passes it through untouched.

Two different problems

Prompt caching makes the same prefix cheap to replay. PromptCrunch makes a growing conversation history smaller before it's sent.

Provider prompt caching

Built into Anthropic and OpenAI. Use it. It's excellent at what it does.
  • Stable prefixes. A large frozen system prompt or document replayed across many requests gets served from cache at a fraction of the input price.
  • Agentic loops. Coding agents and tool-use loops replay the same system prompt and tool definitions every turn. Caching covers exactly that.
  • No middleman. It runs inside your provider. Nothing new in your request path.
  • The catch: prefix match. One changed byte invalidates everything after it, and the conversation still grows. You pay (cached rates) for an ever-longer prompt.

PromptCrunch compression

A proxy that shrinks old conversation history before forwarding.
  • Long prose conversations. Chat products, tutoring bots, coaching apps, support sessions. In our benchmarks: 60-75% fewer input tokens past 20 turns.
  • Stacks without caching. OpenAI-compatible endpoints, self-hosted models, and frameworks where you can't or don't set cache breakpoints.
  • Shrinks, not replays. The forwarded prompt is actually smaller, so it works even when every request has a different prefix.
  • The catch: prose only. Code, JSON, tool output, and recent turns are preserved verbatim, so code-heavy traffic saves almost nothing (0-7% measured).

Our honest recommendation

If your requests already use prompt caching, keep it and skip us. PromptCrunch detects cache_control and passes those requests through completely untouched, breakpoints intact. We never break a cache to force savings, which also means we can't add savings on top of a well-cached agentic workload.

If you run long, prose-heavy conversations without caching, especially on OpenAI-compatible or self-hosted stacks, compression is the tool that actually shrinks the bill. That's the workload we built for, and the only one we'll claim numbers on.

Pick by workload

Caching wins

  • Coding agents (Claude Code, Codex)
  • RAG over a stable document set
  • Large frozen system prompts replayed at volume
  • Tool-heavy agent loops with stable definitions

Compression wins

  • Companion and character chat
  • Tutoring and education bots
  • Coaching apps and long-session support
  • Any 20+ turn prose conversation without cache support

Mixed workload? Route everything through PromptCrunch: cached and code-heavy requests pass through, conversational ones get compressed.

Benchmark it on your own traffic

$5 free credit, 100 requests/day. No card.

Try it free vs native compaction