PromptCrunch vs provider prompt caching

Caching cuts your bill when you replay the same prefix. PromptCrunch cuts it when the conversation grows, up to 75% fewer input tokens on a long chat, with the same replies word for word. Already caching? Your cached requests pass through untouched, so you're never worse than direct. Run both. The savings stack.

Try it free See the two-line setup

Caching and compression cut different bills

Prompt caching makes the same prefix cheap to replay. PromptCrunch shrinks a growing conversation before it ever reaches the provider. Most real products carry both kinds of traffic, so run both.

Provider prompt caching

Built into Anthropic and OpenAI. It does one thing well, replaying a fixed prefix.
  • Stable prefixes. Replay a big frozen system prompt or document across many requests and the provider serves it from cache at a fraction of the input price.
  • Agentic loops. Coding agents and tool-use loops resend the same system prompt and tool definitions every turn, exactly the fixed prefix caching was made for.
  • No middleman. It runs inside your provider. Nothing new in your request path.
  • Then the conversation grows. Cache only helps the part that never changes. Every new turn is fresh, full-price input, and the prompt gets longer with every message. Caching can't shrink what it can't repeat.

PromptCrunch compression

A proxy that shrinks old conversation history before it's forwarded.
  • The bill drops, hard. On a 40-prompt Claude Sonnet run in our benchmarks, the input bill fell from $2.56 to $0.64, a 75% cut with identical responses. That's chat products, tutoring bots, coaching apps, long support sessions.
  • Different prefix every time? Still cheaper. Caching needs an exact repeat to help. PromptCrunch forwards a genuinely smaller prompt, so the bill drops even when no two requests start the same.
  • No breakpoints, no config. Works with Claude and OpenAI, any SDK with a configurable base URL. Two lines, and the savings land on the first request.
  • Nothing that matters gets touched. Code, JSON, schemas, IDs, and your most recent turns pass through verbatim. The model still sees everything it needs, so the replies don't change.

Never worse than direct, often much cheaper

Your cache stays exactly as cheap as it is today. PromptCrunch sees cache_control, leaves those requests untouched with breakpoints intact, and forwards them straight to the provider. We never break a cache hit to manufacture savings, so routing through us is never worse than routing direct.

Everything the cache can't help, we take on. The long, growing conversations, chat, tutoring, coaching, support, where the prompt swells every turn and full-price input piles up. Point that traffic at PromptCrunch and the input bill falls. Cached requests and live conversations, one endpoint, both cheaper.

Match it to your product

Cache carries these

  • 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

PromptCrunch crushes these

  • Companion and character chat
  • Tutoring and education bots
  • Coaching apps and long-session support
  • Any conversation that keeps growing, cached or not

One base URL, one smaller bill. Try it on your own traffic - the token counter tells the truth in minutes.

Point your traffic at it and watch the input bill fall

Free to start with $5 of credit and 100 requests a day, no card. Two lines of setup, then run a real conversation through and watch the token count drop.

Try it free Compare vs native compaction