Questions people ask before switching

No. If every request needs the strongest model available, call that model directly. Auto pays off on long agent sessions where most of the work is routine; Fusion pays off when one answer has to be right and you can wait a little longer for it. Test on your own workload before you switch.

Take a real workload with a clear success criterion: tests that pass, a reviewer who signs off. Run it on your current model and on Modeless, then compare quality, latency and cost per finished task, not per token.

Auto changes model at most once per session, when it decides the task needs a stronger model. That one switch starts a new prompt cache; every turn after it stays on the same model and the cache stays warm. The one exception is a model failing before it answers: that request falls back to another model.

Modeless only moves to another model before the first byte of a reply reaches you. Once a reply has started, it never splices a second model's output into it; you get the error in your protocol's own format instead.

You pay what the models cost, plus 5%. Your console lists every request with the model that served it, its tokens, its unit price and the total.

To the models that serve your request. Auto also sends the task and the agent's working history to a model that judges how hard the task is. Data & Security lists what is stored and for how long.

OpenRouter is a catalog: you pick the model. Modeless gives you two models that make the choice for you: Auto picks the model as a session goes, and Fusion combines several. If you already know the one model you want, a catalog may suit you better.

Both models speak Chat Completions, Anthropic Messages and Responses, streaming or not. Auto only routes a request to models that support what it contains: images, tool calls, thinking parameters, its context length.

No. When your agent calls tools several times inside one turn, the panel's answers from the first call are reused. You pay for the panel once per user message, not once per tool call.

The aggregator works with the answers it has. A Fusion request fails only if the aggregator itself fails.