vrinda.dev / compare
the race lane — same prompt, different discipline

same prompt · same key · 2-3 configs

race two orchestrations.

Give one prompt to two ways of building it. Watch each config run live on your own OpenRouter key, then judge them side by side — metered cost, wall time, and the artifacts themselves. The exact runner behind every comparison on this site.

up to 3

This exact prompt goes to every lane. 8 KB (8192 bytes) cap — UTF-8 counted.

try:

your key stays in this browser — we never store it

your key goes directly to OpenRouter through each run and is never stored on our server — receipts keep an 8-char fingerprint only

Kept only in this browser's localStorage so you don't retype it. Clear it anytime with the button below. One key runs every lane.

actual cost is metered from provider receipts · coarse server ceilings apply (currently $1.00/run · $5.00 rolling session) · public-by-link: a shared board shows your prompt, configs, and artifacts

recorded races — receipts on the board; no models are running here.

saved — my races & configs

how this works — and its limits

Bring-your-own-key (BYOK). No signup, no account, no balance held.

  • Your key is yours. At race time it travels browser → our runner → OpenRouter and is never written to our queue, database, logs, or receipts. Receipts keep only sha256(key)[:8] so you can match a run to your key without us holding it. The opt-in coach is the separate browser → OpenRouter direct path and never touches our server.
  • You pay OpenRouter, we don't. Every dollar figure is metered from the provider's own generation receipts and billed to your OpenRouter account, not ours. Estimates are used only for a pre-call budget reservation.
  • A race is 2-3 real runs, executed one at a time by a single worker — lanes queue in turn; a lane can fail alone without killing the race. Per-run and rolling session ceilings refuse runaway spend; a wall-clock deadline and a rolling per-address allowance (currently 12 runs/hour) apply.
  • Share = public. The compare link reveals your prompt, the lane configs, receipts, and the artifacts to anyone who has it. Don't race secrets.
  • Featuring is curated. Machines may shortlist candidates — always labeled as machine-shortlisted — but a human makes every quality call. No machine ranks or crowns quality here; on your race board, quality stays your call.
  • The artifacts are real files. Each final artifact.html is served from a locked-down origin under a strict sandbox CSP. It is unlisted (a random run id) but not access-controlled — treat anything you build as shareable.
  • Free routes are rate-limited. free routes are rate-limited by OpenRouter (~50 requests/day on a $0 account) and may log prompts — don't race secrets on free models.

the sealed harness case files › — the same runner, run under a fixed harness on a standing matrix.