Paste a transcript. See what we'd really save.

This estimator runs the same decision rules as the live proxy: the 12,000-token floor, the keep-the-last-4-messages-verbatim rule, and verbatim preservation of code and structured data. If your traffic is a bad fit, it says so.

Runs entirely in your browser. Your transcript never leaves this page.
Paste
A real conversation
Simulate
Same rules as the proxy
Read
An honest verdict

Already using prompt caching (cache_control)?

Those requests pass through untouched - model your savings as zero and keep your cache. We never break a cache to force savings. Full comparison here.

What this simulation actually does

No black box. These are the same rules the live proxy applies, minus the summarization model itself.

  1. Token estimate: characters ÷ 4. Crude but consistent; real tokenizers land within about 10% on English prose. Base64 image blocks count as a flat ~1,500 tokens each.
  2. Passthrough floor: if the whole conversation is at or under 12,000 tokens, the proxy forwards it untouched. You pay nothing extra and save nothing.
  3. Recent turns: the last 4 messages are always forwarded verbatim. The model needs full fidelity on what was just said.
  4. Protected content: fenced code blocks, JSON-looking messages (starting with { or [), base64 runs, images, and structured blocks (tool calls, tool results) in older history are preserved verbatim, never compressed.
  5. Everything else in older history is compressed at a 0.4 target ratio: forwarded = protected + 0.4 × compressible + last 4 messages.
  6. Honesty gate: if protected content is more than 60% of the older history, we tell you the savings are not worth it, because for that traffic they usually are not.

Numbers look good? Test them for real.

The free tier's $5 credit against your live traffic beats any simulation.

Try it free ($5 credit, no card) PromptCrunch vs prompt caching