Open it from Create → Model Test, or from the link on the Generate and Tools pages.
Pick the engines you want to try, write one prompt, and every selected engine runs it. You compare the results in a grid and pick one to move forward with.
What makes the comparison honest: every entry gets the same prompt, same seed, same references and same duration. A bake-off where each model got a different seed is a comparison of luck, not of models — and acting on that could send you down a six-week production path on noise.
Engines that cannot run are shown, not hidden. Each one names the exact key it needs. A grid of four results when you asked for seven is misleading unless it says what happened to the other three.
Picking a winner does something. It becomes the project's preferred engine for that surface, and every later generation on the project uses it unless a shot overrides it. Without that step this would be a nice picture grid that changes nothing.
How it differs from arena mode: arena compares models for a single shot mid-edit, and the Router picks the candidates. This is an up-front decision about the whole project, and you pick the candidates — including ones the Router would never choose on its own.
A pinned model never falls back. Normally a provider outage routes to the next-best engine. During a test that would be wrong: you asked what a specific engine does with this prompt, so if it cannot run, the entry fails and says so rather than quietly showing you a different model's work.
With no provider keys you get one result from the local mock adapter plus a card per engine naming the key it needs. That is not a real comparison and the page says so — but it shows you exactly what the feature does and what it would cost to make it real.
Cost is shown before anything runs, and each result reports its actual credits, latency and quality score so you can weigh look against price.