Structured, step-by-step thinking on demand. Pick the Reasoning interceptor from the sidebar when you want the model to slow down and show its work.
Reasoning is one of three interceptors Berges AI ships with. Pick it from the sidebar and the deep-thinking layer switches into a mode tuned for structured, multi-step thought: lay out the inputs, work through them in order, land on a conclusion.
It's the right choice when the answer matters more than the prose, and when getting to the answer requires staying organized through more than one or two hops.
Reasoning uses a low sampling temperature, which keeps the model consistent across runs. Ask the same question twice and you should get a similar shape of answer, which is what you want for analysis.
On top of whatever model is running, Reasoning layers a system prompt that asks for steps, evidence, and a conclusion. The Berges-style framework is the proprietary part.
Reasoning only kicks in on escalated turns. Easy questions still go through the fast layer untouched, so you don't pay the cost of structured thinking on a trivial question.
When the question is easy or open-ended. Reasoning makes the model slower and more methodical, which is the wrong shape for "give me five name ideas" or "what time is it in Tokyo".
Related but not identical. Saying "think hard" (or "think harder") in a message forces escalation to the deep-thinking layer for that turn. Reasoning, when selected, also escalates and additionally applies the structured-thinking system prompt.
Yes. Interceptors operate on top of whichever model the cascade picks, so the Reasoning system prompt and lower temperature apply regardless of whether the underlying model is Kimi, DeepSeek-Pro, or MiniMax.