Run it on your own traffic.

Create an account and you get an API key straight away. Point your existing client at the engine, keep your pinned model, and it records what it would have chosen instead. Nothing changes until you decide it should. If your security review needs it, run the shadow on synthetic traffic and a test provider key first; a production key is never required to see the result.

• An API keyPoint your existing client at the engine. One environment variable, your code unchanged.
• Shadow modeIt serves your pinned model untouched and records what it would have chosen. No spend, no risk.
• Your own numbersThe counterfactual on your traffic — including when the answer is that we would not win.

1.856–1.926 µs per request served · 3.04–3.20 million a second across one server
GOOGLE CLOUD C3-STANDARD-8 · MEASURED ACROSS MULTIPLE RUNS