> A common misconception is that AI is more expensive than search. Opposite is true - performing a search is 100x-1000x more expensive than doing an interaction with AI.
> A single search is about ~ 1.5 cents (probably bit less these days, but general ballpark)
> A small AI model can generate a lot of tokens for the same price
> To put this in perspective, 1.5 cents of Gemini Flash usage:
> 50,000 input tokens ≈ 37,500 words (roughly 75-150 pages of text)
or
6,000 output tokens ≈ 4,500 words (roughly 9-18 pages of text)
> That is A LOT of information for the cost of just 1 search.
> It may be counter-intuitive but that is how it is. If someone is using AI to answer common questions, learn about a concept, do a quick brainstorm or a translation - instead of searching - our cost is (much) lower.
> So if anything, having AI lowers our cost, not increases it. This is why when we added access to AI models to Pro tier we didn't increase the price.
Doesn't make sense because AI responses are not grounded, still for AI to make sense in this context, and have any relation with Kagi purpose, you need to have the search still, and then the AI process the search results.
https://kagifeedback.org/d/1338-provide-a-plan-without-ai-fu...
> A common misconception is that AI is more expensive than search. Opposite is true - performing a search is 100x-1000x more expensive than doing an interaction with AI.
> A single search is about ~ 1.5 cents (probably bit less these days, but general ballpark)
> A small AI model can generate a lot of tokens for the same price
> To put this in perspective, 1.5 cents of Gemini Flash usage:
> 50,000 input tokens ≈ 37,500 words (roughly 75-150 pages of text) or 6,000 output tokens ≈ 4,500 words (roughly 9-18 pages of text)
> That is A LOT of information for the cost of just 1 search.
> It may be counter-intuitive but that is how it is. If someone is using AI to answer common questions, learn about a concept, do a quick brainstorm or a translation - instead of searching - our cost is (much) lower.
> So if anything, having AI lowers our cost, not increases it. This is why when we added access to AI models to Pro tier we didn't increase the price.
Doesn't make sense because AI responses are not grounded, still for AI to make sense in this context, and have any relation with Kagi purpose, you need to have the search still, and then the AI process the search results.
Common facts like “what is the capital of Hungary” are repeated so many times in the training data that the LLM knows them without a search.
Makes sense.
Thank you.