Instead of one big model for everything, a cascade starts small and escalates. Here is how the routing decision is made and why it changes the economics of AI chat.
Most AI products pick one model and send every message to it. That is simple, and it is wasteful in both directions: a huge model answering "what time is it in Tokyo" burns money and seconds for nothing, and a small model asked to plan a migration gives a shallow answer with full confidence.
A model cascade is the alternative. Each message goes to a small, fast model first. If that model judges the question to be beyond it, it escalates, and a larger reasoning model takes over. The user sees one conversation. Underneath, two or more models are sharing the work.
There are two common designs. One puts a separate classifier in front that scores each message for difficulty. The other asks the small model itself to decide: answer if you can, or signal that this needs the bigger model. The second is simpler and uses what the small model already knows about its own limits, which is more than you might expect. Either way, the rule of thumb is that borderline cases escalate, because a wrong "easy" call is costlier than a wasted escalation.
Typical escalation triggers: the question needs several dependent steps, it needs specialized knowledge the small model is unsure of, it asks for careful synthesis across a lot of material, or the user explicitly asks for more effort. Short factual questions, rewrites, quick translations, and casual conversation stay on the small model.
In a typical chat mix, most messages are easy. If the small model handles those, the large model runs on a fraction of the traffic, and that fraction is where it earns its cost. The result is faster average responses and a much lower bill, which is what makes serving large open-weights models at everyday prices possible.
The failure mode is a confident small model answering a hard question without escalating. Every cascade leaks some of these. The fixes are a bias toward escalating when in doubt, a way for the user to force the larger model, and transparency about which models sit in each layer so the user can judge. A cascade that hides its routing is harder to trust than one that shows it.
Cascades became more valuable once reasoning models arrived, because those models are slow and expensive by design. Sending everything to one would be painful. Sending only the hard questions to one is close to ideal: the thinking cost lands exactly where thinking helps.
It is how a good clinic works. A nurse handles most visits and knows which cases need the doctor. The doctor sees a fraction of patients, the ones who need that time, and the whole clinic runs faster and cheaper than one where every visit starts with the specialist.
Berges AI does not cascade. Every turn goes to the reasoning model you pick from the sidebar, with no router deciding you asked an easy question. The gap a cascade exists to bridge has mostly closed: quick, cheap reasoning models such as DeepSeek Flash think through a routine question in about the time a plain model takes to answer it, so a small model in front saves little and costs you the working whenever the router gets it wrong.
Try Berges AIBecause every model on Berges AI already reasons, and you choose which one answers. A router in front would second-guess that choice and would sometimes keep a hard question away from the model it needs. Sending every turn to the model you picked is simpler and easier to trust.
On Berges AI, yes. You pick the model from the sidebar and it answers every turn. In general it depends on the product. If one does not tell you how it routes, assume the routing is tuned for the provider's cost, not your question.
No. A mixture of experts is inside one model: it routes each token to parts of the network. A cascade routes whole messages between separate models. They can coexist. Every model on Berges AI is a mixture of experts, and none of them sits in a cascade.
Per reply, yes, because it thinks before it writes. The gap has shrunk a lot: the smaller reasoning models handle a routine question in a few seconds, and on a hard one the extra time is where the quality comes from.