The Kagi Research assistant is fantastic, but it's pricy. Lately I've been using a lot a custom assistant with the following prompt:
Search the web for as much information about the topic as possible. Make multiple searches with different keywords whenever it makes sense. If the user's question is asking for a recommendation, choose a recommendation, but list potential drawbacks, along with potential advantages or situations where you'd want the other choice.
With gemini 3.1 flash lite, it is imo better than the kagi quick assistant and about as fast. With Kimi K2.5 reasoning, it is slower but significantly better. It's still significantly worse than the research assistant, but it's cheap enough I can use it for anything, while the research assistant is pricy enough to only be able to use a few times a day. I think it'd be a nice option to be able to use the research assistant with a cheaper model.
Currently I think the research assistant mostly uses sonnet? IMO kimi k2.5 and glm5, and probably deepseek 4 when it comes out, are close to sonnet and much cheaper. The research assistant has access to more tools and a whole agent orchestration designed for research so I assume those models driving the research agent will be significantly better than the same model with just web search. It could even be a competitive differentiator for kagi assistant, if the "lite" research agent with cheaper models performed well but could be used more often than the "deep research" agents on perplexity, google, chatgpt, or other subscriptions.
