Dynamic Kagi Recommended/SOTA Default Assistant

CrunchyFritos

I am suggesting the addition of an option to set a "Kagi Recommended" model as a user's "default assistant" where the Kagi Recommended model is automatically the best model available for that user based on (1) that user's subscription plan and (2) the Kagi LLM benchmark, and (3) Kagi staff's qualitative judgement.

I see this feature implemented in two parts:

Adding a "Kagi Recommended" default assistant option that automatically selects the best non-thinking / auto-routing model available for the user based on that user's subscription level.
Adding a !best bang that automatically selects the absolute best SOTA model available to the user with internet turned off (if I wanted internet on I would likely just use the new Kagi Research Assistant).

With this feature, I could just "set it and forget it" with respect to my Kagi assistant settings, knowing that if I use the !chat and !best bangs, I will always be getting pure, unadulterated access to what is likely the best model available to me at that time.

fxgn

The two Kagi assistants (previously named Ki, now just called Quick and Research) are supposed to be exactly that. The Quick assistant is what the Kagi team chose as the experience with the best balance of speed and intelligence for quick queries that don't require deep research.

CrunchyFritos

fxgn Not really, the Kagi assistants seem to default to some version of Claude Haiku as their base model. So if you trigger them with a bang or as a default model without web access you don’t really get the “best” model available to you based on your subscription (for example Claude 4 or GPT 5 if I’m an Ultimate Plan subscriber). The feature I’m suggesting is a default model that will automatically select the best model available to you, and then update that underlying model as the Kagi team changes the menu of models available to me or as the LLM benchmark rankings change.

Thibaultmol

best is relative though.
The objectively best model right now is Opus but it'll also bankrupt you if you were to use it consistently (if you get what I mean).

Quick and Research are indeed supposed to be the defaults that Kagi is recommending to people.

As Matt (ai lead at kagi) mentioned to me, it's not just about the model but also about all the other settings and the prompt. But yeah, they recommend Quick and Research atm

I do wonder if this is possible: have a profile name that just redirects, so you could have profile name like kagi-recommended which you could use to create a bang like !airec
I doubt they'd make a bang like that which has internet off BUT I think they could then make !chatrec ( !chat is the same as !ai but without internet access).
And so that would just be redirecting to whatever is at the top of the 'Kagi Recommended' section.
I'll ask

MattM

Right, so the "best" model is dependent on a bunch of things:

What task are you doing? For a lot of tasks the best model is the cheapest & fastest because the task is easy. For coding, the best model is different than for in depth research.
The "best model" is generally more the best "experience", as @Thibaultmol said -- so the integration of the whole flow from query to answer is what people care about and models behave differently. So the Kagi Assistants is the best way to resolve that by having a setup answer.
For a bang to a specific model, you can do it with a bang to the model id (eg. !kimi-k2). For a specific experience you want, I'd say the best experience is making a custom assistant and tying a bang to it? You could add slash commands in the custom prompt to ensure web acces or not, etc.