live demo

Only the hard part goes to Claude.

Paste a coding prompt. Jev decides which parts are routine, a cheap model answers those, and you see what Claude would have charged for them.

0 / 600
Try:

How this demo works

Your prompt is split into clauses. Jev, a fast judgment model from TypeSafe on OpenRouter, rates each clause as routine or hard. Routine clauses go to a short cascade of cheap models; if one fails or is rate-limited, the next one answers. Hard clauses would go to Claude, so nothing is sent for them here.

Savings are an estimate at Claude Sonnet list prices ($3 input and $15 output per million tokens) for the prompt Claude no longer reads and the answer it no longer writes, minus what Jev and the cheap model actually cost. Per prompt it is small; it adds up across thousands of prompts.

Limits keep this page cheap to run: 600 characters per prompt, a few tries per minute, and short answers. Prompts are sent to OpenRouter to be answered and are not stored by waterfall.sh. Don't paste secrets or private code.