Hi community! I've been on board with Kagi for some months, and the general experience is really nice (coming from Searx, Kagi can provide the same or better results w/o interruptions)! Also, since I've been using some LLMs in my life and work recently, I've used Kagi assistant a few times and it's also been great.
However, seeing the more advanced models in the Ultimate plan, I still cannot convince myself to go for that mainly due to pricing issues:
Sure, 25 bucks altogether is a bargain when considering subscribing ChatGPT Plus (or Google AI Pro, or Claude Pro) and Kagi costs 30, but here are the caveats:
- I'm not a die-hard privacy lover, so I do appreciate some free and nice options out of there (for example, Google AI Studio and Gemini CLI can give me A LOT of usages of Gemini Pro free);
1.1 Therefore, if I decide to return to Searx with free usages of Gemini Pro, it's a totally free business (again I'm not a heavy lover of privacy), and costs me $5 less if I decide to subscribe to ChatGPT Plus or Perplexity Pro, etc.;
- Models like Codex are mostly reported helpful in coding, but for me Kimi K2 and GLM 4.6 are really enough in my scenarios, and they are really cheap by API usages or even monthly subs in my country (like a world class bargain of 50%+ compared to these international models);
- Kagi's pricing model is a bit strange: I've got a $10 sub, but my LLM usage is calculated by usages, and it will block further usage when these $10 are used. I know Kagi is not openAI or Google and paying API fees would cost a lot, but it still limits the potential and my willings to using Kagi Assistant since the large players can simply limit you to lesser performant models instead of blocking everything. And the LLM models unlocked in the Ultimate plan can easily eat up all the 25 bucks.
It's really hard to make a decision on the matter, but what if the stair model can be adopted here on Kagi? Let's say (only my personal thoughts) if a user uses too much of a high-performance and high-price model, we can limit him to lower ones, and if he still uses them up, we can limit him to some in-house-deployed lightweight models. This model could keep the access to Assistant at all times.
For example:
One used too much of GPT-5 -> limit to GPT-5 mini or nano -> mini or nano were also used up -> limit to let's say, GPT-OSS-20B -> used up again -> limit to Qwen3-8B (or some lightweight competitors)
(BTW, we can get an AI-free or free usage of Assistant with these super lightweight models)
Still, I understand your guys' business decisions and I am content of my current plan (so no complain if we have to stick to the current divisions of subscriptions). Hope you have great days and keep the momentum of improving Kagi nicely!
