To summarize, Kagi's LLM interface is available on all their plans, with usage limits depending on the plan, and the more costly models only available for "Ultimate" subscribers.
To summarize, Kagi's LLM interface is available on all their plans, with usage limits depending on the plan, and the more costly models only available for "Ultimate" subscribers.
Which sounds fair enough (as a regular non-ultimate subscriber). Although I can’t imagine they’re expecting the margins to work super well for them unless most people don’t use most of the...
Which sounds fair enough (as a regular non-ultimate subscriber). Although I can’t imagine they’re expecting the margins to work super well for them unless most people don’t use most of the allotted tokens most of the time.
At least they now seemingly have a mechanism to enforce the usage limits, the stat about ten users (not percent) causing 14% of total LLM cost is nuts.
I always find numbers like this fascinatingly unintuitive - I think we tend to assume our own use of whatever service is somewhere nearish the average; maybe up to double if we consider ourselves...
I always find numbers like this fascinatingly unintuitive - I think we tend to assume our own use of whatever service is somewhere nearish the average; maybe up to double if we consider ourselves a heavy user, down to half as a light user, but all within the same ballpark. Actually looking at the data rarely matches up to that, in my experience!
I distinctly remember a while back when they published search volume statistics and it turned out my fairly standard Tuesday was above the average monthly usage, for example. Equally, I don’t use their AI tools at all, so that’s at least one user’s worth of tokens back in the pool to pad the margins a bit.
I do wonder about those ten people using 50M tokens each though! Part of me thinks they’ve got to be deliberately doing something “unreasonable” specifically because there’s an unlimited resource, but equally I guess if you find it does a good job of, say, parsing entire GitHub repos to find bugs maybe you could hit those levels without intentionally doing anything wildly out of scope for the tool.
Plus, they'll allow you to buy "credits" in case you need more than it'll be defined in their fair usage policy. So you're not blocked by their limits. While I love Kagi search and the quick...
Plus, they'll allow you to buy "credits" in case you need more than it'll be defined in their fair usage policy. So you're not blocked by their limits.
While I love Kagi search and the quick answer feature, after trying many times the Kagi Assistant (with a temporary "Ultimate" subscription) I always felt like it is lacking with the answers I've gotten.
Being a Perplexity Pro user for a long time, I've received rather accurate and detailed answers whenever I used it. Trying out the paid ChatGPT version or Kagi Assistant felt like a downgrade, and asking similar questions I could often see in direct comparison that the answers were either less detailed or not accurate.
I wonder why, not sure what Perplexity is doing differently. Maybe their sources they search for or that they almost always do a web search, even if it seems like the question asked wouldn't need any web results?
Hope to see Kagi Assistant improve in the future. Kagi Translator is already top-notch and is giving better translations than I get with DeepL.
Outside of looking up local businesses, I haven't found Google any better. Kagi's results are partially from Google's own index, so I'm a bit puzzled. Since Kagi lets you apply custom weights and...
Outside of looking up local businesses, I haven't found Google any better. Kagi's results are partially from Google's own index, so I'm a bit puzzled. Since Kagi lets you apply custom weights and rankings to websites, I've found it much better with my customisations.
While it's neat that Kagi is bringing this to their lower tier plans, I have to admit I'm a little annoyed that they won't decrease the cost of the unlimited search plan. Clearly there's overhead...
While it's neat that Kagi is bringing this to their lower tier plans, I have to admit I'm a little annoyed that they won't decrease the cost of the unlimited search plan. Clearly there's overhead in there for AI credits; could I just pay $5/mo for ONLY unlimited search instead?
It doesn't break the bank or truly matter for me, but it's a huge problem when I try to convert friends from Google, DDG, or Yandex.
I had a similar reaction. As someone with zero interest in AI tools, my main concern with Kagi is the price. I renewed it for a second year, but it was a tough decision, and I'm not sure I can...
I had a similar reaction. As someone with zero interest in AI tools, my main concern with Kagi is the price. I renewed it for a second year, but it was a tough decision, and I'm not sure I can justify it indefinitely as a casual web browser who isn't doing heavy research.
I very recently switched up to Ultimate to try out the assistant, and it’s been… fine. I think the interface is honestly kind of janky, and inferior to something like ChatGPT’s. The ability to...
I very recently switched up to Ultimate to try out the assistant, and it’s been… fine. I think the interface is honestly kind of janky, and inferior to something like ChatGPT’s. The ability to switch between different LLMs is useful, but I simply don’t use it. I’m not sufficiently clued up on the tech to know which LLM is the right one for my needs.
Anyway, this is good news, and I’m dropping my sub back down to the Pro one, and saving myself a healthy sum in the process. I don’t know if I’ll continue using Kagi assistant as my go-to LLM though. Maybe I’ll put the money saved towards a different LLM sub.
Not sure how they can be faster considering they will be using the same API endpoints. If you look at the FAQ it becomes more clear, they basically are saying they are better at making fast...
Not sure how they can be faster considering they will be using the same API endpoints. If you look at the FAQ it becomes more clear, they basically are saying they are better at making fast "feeling" application frontends.
But again, given they are using the same APIs the information isn't going to be actually being processed faster, it might just feel faster compared to some other chat front ends.
To me it seems like just another LLM chat frontend.
So, update, I guess? Your point got me thinking, since, yeah, what exactly is the value proposition when “fast” has a lower bound it cannot ever surpass? (Again, maybe I should’ve pointed out I...
So, update, I guess?
Your point got me thinking, since, yeah, what exactly is the value proposition when “fast” has a lower bound it cannot ever surpass? (Again, maybe I should’ve pointed out I had only talked with others about the thing I was seemingly recommending, not used it before myself.)
The answers are in the developer-turned-content-creator’s “AI chat apps are driving me insane” (warning: 1+ hour rant video of a, let’s say, dedicated engineer).
I’ve only begun watching it, but think I can already tell a pattern: if you’re the sort of person to be annoyed by bad (web) design, poor implementations, lacking UI – all these smaller nitpicks you may find in other providers’ first-party applications – summed up together, you may see enough value in UX improvements alone that a subscription to this instead becomes worthwhile.
Edit: This is a worthwhile snippet at 33:36 onward. Even in the video, I as a not overly technical viewer of a screen recording can tell which of the two is more performant as well as correct (!) in comparison. And it’s literally the first-party app he’s testing against.
Edit 2: They also have to work around quirks found in the models’ APIs…
Same. I should’ve probably disclaimed that I’m not actually a user myself, I’ve been happy speaking to the various models via smaller “direct” API clients so far.
Same. I should’ve probably disclaimed that I’m not actually a user myself, I’ve been happy speaking to the various models via smaller “direct” API clients so far.
I recently prorated the last few months of my professional sub up to ultimate so that I could try the assistant out. I've found it useful, but I think anyone who's used to the multimodal...
I recently prorated the last few months of my professional sub up to ultimate so that I could try the assistant out. I've found it useful, but I think anyone who's used to the multimodal capabilities of ChatGPT would be underwhelmed by it. The ability to generate and upload images and process audio and video seems like it would be much more useful to the average user than being able to switch between different LLM models (which is probably Kagi's differentiating feature). Now that it's available on professional plans I might switch back when it comes time to renew, because I'm not sure whether my limited use cases even benefit from the fancier models.
⚠️ Note: We are enabling the Assistant for all plans in phases, based on regions starting with USA today. The full rollout is scheduled to be completed by Sunday, 23:59 UTC.
Huh. That's interesting. I tried out Kagi for a while but the search results were not worth $10 a month in my opinion. But I have been thinking of paying for AI to help me debug some of my code,...
Huh. That's interesting. I tried out Kagi for a while but the search results were not worth $10 a month in my opinion. But I have been thinking of paying for AI to help me debug some of my code, so this might just make it worth the subscription price even if it's limited. But honestly, right now @tauon's suggestion of T3 seems like it wins out because it's nearly half the price. Though it's a bad sign when I asked the FAQ if their price included unlimited usage it said to contact them directly... When I asked how to do that it said to try a whois lookup.
Edit: Actually, it looks like the only way to sign up is with a Google account, so I'll probably pass on them.
For coding, I can recommend Github Copilot or, if you rather script, checking out Warp. When compared to other AI tools, these integrate well into workflows and are better when providing...
For coding, I can recommend Github Copilot or, if you rather script, checking out Warp.
When compared to other AI tools, these integrate well into workflows and are better when providing suggestions, helping debug code or even writing code (snippets) for you.
I'd also suggest trying something like Cursor depending on what you want out of the tooling. It's very expensive though, so it's only worth it if work is paying for the LLM in my opinion.
I'd also suggest trying something like Cursor depending on what you want out of the tooling. It's very expensive though, so it's only worth it if work is paying for the LLM in my opinion.
Yeah, it was a secondhand suggestion, not from my personal experience, since I got stopped at the “only Google” sign-up too. I really wonder why they haven’t added more auth options yet.
Yeah, it was a secondhand suggestion, not from my personal experience, since I got stopped at the “only Google” sign-up too. I really wonder why they haven’t added more auth options yet.
To summarize, Kagi's LLM interface is available on all their plans, with usage limits depending on the plan, and the more costly models only available for "Ultimate" subscribers.
Which sounds fair enough (as a regular non-ultimate subscriber). Although I can’t imagine they’re expecting the margins to work super well for them unless most people don’t use most of the allotted tokens most of the time.
At least they now seemingly have a mechanism to enforce the usage limits, the stat about ten users (not percent) causing 14% of total LLM cost is nuts.
I always find numbers like this fascinatingly unintuitive - I think we tend to assume our own use of whatever service is somewhere nearish the average; maybe up to double if we consider ourselves a heavy user, down to half as a light user, but all within the same ballpark. Actually looking at the data rarely matches up to that, in my experience!
I distinctly remember a while back when they published search volume statistics and it turned out my fairly standard Tuesday was above the average monthly usage, for example. Equally, I don’t use their AI tools at all, so that’s at least one user’s worth of tokens back in the pool to pad the margins a bit.
I do wonder about those ten people using 50M tokens each though! Part of me thinks they’ve got to be deliberately doing something “unreasonable” specifically because there’s an unlimited resource, but equally I guess if you find it does a good job of, say, parsing entire GitHub repos to find bugs maybe you could hit those levels without intentionally doing anything wildly out of scope for the tool.
Plus, they'll allow you to buy "credits" in case you need more than it'll be defined in their fair usage policy. So you're not blocked by their limits.
While I love Kagi search and the quick answer feature, after trying many times the Kagi Assistant (with a temporary "Ultimate" subscription) I always felt like it is lacking with the answers I've gotten.
Being a Perplexity Pro user for a long time, I've received rather accurate and detailed answers whenever I used it. Trying out the paid ChatGPT version or Kagi Assistant felt like a downgrade, and asking similar questions I could often see in direct comparison that the answers were either less detailed or not accurate.
I wonder why, not sure what Perplexity is doing differently. Maybe their sources they search for or that they almost always do a web search, even if it seems like the question asked wouldn't need any web results?
Hope to see Kagi Assistant improve in the future. Kagi Translator is already top-notch and is giving better translations than I get with DeepL.
I've tried Kagi, but google is still leaps and bounds ahead
Even Google's AI mode now trounces perplexity
Outside of looking up local businesses, I haven't found Google any better. Kagi's results are partially from Google's own index, so I'm a bit puzzled. Since Kagi lets you apply custom weights and rankings to websites, I've found it much better with my customisations.
What kind of stuff do you usually look up?
While it's neat that Kagi is bringing this to their lower tier plans, I have to admit I'm a little annoyed that they won't decrease the cost of the unlimited search plan. Clearly there's overhead in there for AI credits; could I just pay $5/mo for ONLY unlimited search instead?
It doesn't break the bank or truly matter for me, but it's a huge problem when I try to convert friends from Google, DDG, or Yandex.
I had a similar reaction. As someone with zero interest in AI tools, my main concern with Kagi is the price. I renewed it for a second year, but it was a tough decision, and I'm not sure I can justify it indefinitely as a casual web browser who isn't doing heavy research.
I very recently switched up to Ultimate to try out the assistant, and it’s been… fine. I think the interface is honestly kind of janky, and inferior to something like ChatGPT’s. The ability to switch between different LLMs is useful, but I simply don’t use it. I’m not sufficiently clued up on the tech to know which LLM is the right one for my needs.
Anyway, this is good news, and I’m dropping my sub back down to the Pro one, and saving myself a healthy sum in the process. I don’t know if I’ll continue using Kagi assistant as my go-to LLM though. Maybe I’ll put the money saved towards a different LLM sub.
I’ve been told Theo Browne’s T3 Chat is apparently good. As in fast.
Not sure how they can be faster considering they will be using the same API endpoints. If you look at the FAQ it becomes more clear, they basically are saying they are better at making fast "feeling" application frontends.
But again, given they are using the same APIs the information isn't going to be actually being processed faster, it might just feel faster compared to some other chat front ends.
To me it seems like just another LLM chat frontend.
So, update, I guess?
Your point got me thinking, since, yeah, what exactly is the value proposition when “fast” has a lower bound it cannot ever surpass? (Again, maybe I should’ve pointed out I had only talked with others about the thing I was seemingly recommending, not used it before myself.)
The answers are in the developer-turned-content-creator’s “AI chat apps are driving me insane” (warning: 1+ hour
rantvideo of a, let’s say, dedicated engineer).I’ve only begun watching it, but think I can already tell a pattern: if you’re the sort of person to be annoyed by bad (web) design, poor implementations, lacking UI – all these smaller nitpicks you may find in other providers’ first-party applications – summed up together, you may see enough value in UX improvements alone that a subscription to this instead becomes worthwhile.
Edit: This is a worthwhile snippet at 33:36 onward. Even in the video, I as a not overly technical viewer of a screen recording can tell which of the two is more performant as well as correct (!) in comparison. And it’s literally the first-party app he’s testing against.
Edit 2: They also have to work around quirks found in the models’ APIs…
I don’t like that you can only create an account linked with your Google account though.
Same. I should’ve probably disclaimed that I’m not actually a user myself, I’ve been happy speaking to the various models via smaller “direct” API clients so far.
I recently prorated the last few months of my professional sub up to ultimate so that I could try the assistant out. I've found it useful, but I think anyone who's used to the multimodal capabilities of ChatGPT would be underwhelmed by it. The ability to generate and upload images and process audio and video seems like it would be much more useful to the average user than being able to switch between different LLM models (which is probably Kagi's differentiating feature). Now that it's available on professional plans I might switch back when it comes time to renew, because I'm not sure whether my limited use cases even benefit from the fancier models.
:(
Huh. That's interesting. I tried out Kagi for a while but the search results were not worth $10 a month in my opinion. But I have been thinking of paying for AI to help me debug some of my code, so this might just make it worth the subscription price even if it's limited. But honestly, right now @tauon's suggestion of T3 seems like it wins out because it's nearly half the price. Though it's a bad sign when I asked the FAQ if their price included unlimited usage it said to contact them directly... When I asked how to do that it said to try a whois lookup.
Edit: Actually, it looks like the only way to sign up is with a Google account, so I'll probably pass on them.
For coding, I can recommend Github Copilot or, if you rather script, checking out Warp.
When compared to other AI tools, these integrate well into workflows and are better when providing suggestions, helping debug code or even writing code (snippets) for you.
I'd also suggest trying something like Cursor depending on what you want out of the tooling. It's very expensive though, so it's only worth it if work is paying for the LLM in my opinion.
Yeah, it was a secondhand suggestion, not from my personal experience, since I got stopped at the “only Google” sign-up too. I really wonder why they haven’t added more auth options yet.