S2. Ep28 - China's New AI Models Are Nearly as Good as Fable 5 — But Should Your Business Use Them?

Katie (00:26)

Hello, welcome back to another episode. Hi, hello, I'm Katie, and as always I've got Noel here with me today. Hi Noel, how are you doing?

Noel (00:35)

Yeah, doing great as always. How are you doing?

Katie (00:38)

Yeah, good, thank you. So Noel, we always start the podcast off with any new updates with AI and automations, whether it's software or new models.

Katie (00:55)

What are the updates this week? Because last week we spent a whole episode on updates that OpenAI had brought out.

Noel (01:06)

They had loads, didn't they? And it's almost a bit of a similar week this week, but we'll get into a bit more detail on those other bits later on. So Anthropic this week, they decided to end the Fable 5 saga, where they were constantly teasing us, saying, we're going to give you another week, and then we're going to give you another week, and we're like, oh, are you actually going to take it away? And eventually they came to the decision.

Noel (01:35)

For most people, yes, they're going to take it away. So you can only now access Fable 5 if you're on a Teams or Max plan — you've got to be a very high paying user. And even then, you're not actually getting that much usage. So yesterday I hit the usage limit twice. I feel as though the limit has actually gone down a little bit.

Katie (02:02)

Sounds like it.

Noel (02:04)

Probably hasn't — it's probably just how I perceive it, I guess, and what I was doing yesterday. But I just seem to hit it quite quickly. But yeah, it is available now on those plans, and there's no plan yet of bringing it down to the Pro or free users or anything like that. I think one thing that's hampering them is their data centres. They just don't have the capacity right now to give it to everybody. That's kind of their big problem at the moment.

Noel (02:33)

But yeah, if you're on a high pay plan, then Fable 5 is there and ready to go. I'm glad to have put that one to bed — it's been weeks!

Katie (02:40)

How exciting. Yeah, it has. It's been a long time. Just out of interest, Noel, are you willing to tell people what you were doing yesterday and how you hit your usage limit so fast? Because I think that would be actually quite interesting for people to know.

Noel (02:58)

Yeah, I was doing a lot of multitasking yesterday. So I had several coding programs going at the same time using Fable 5. I was doing bits on Clyde, the AI agent platform that I built, and then I was also working on another app with two other guys. So I was doing the coding for their app, and then also their website to go with it. So yeah, I was chewing through the usage quite quickly, unfortunately.

But what I find with Fable is I think it wastes quite a bit of time. It does take a long time to get responses back. So I don't know if it's kind of not that well tuned, I guess is a way to put it. It's taking so much longer, so many more tokens, to get to an endpoint. It's good when you get there, but it's getting through that usage really quickly, unfortunately.

Katie (04:00)

Yeah. And then what would be an alternative to use, then, if you wanted to do those tasks but you don't want to use Fable 5?

Noel (04:18)

So the next best for those sorts of tasks, I think, is currently GPT 5.6 Sol. So on OpenAI, they do have usage limits, but they're not as strict. With Anthropic, you've got a five hourly rolling limit and a weekly limit, whereas I think OpenAI have gone straight for a monthly limit. So if you've got a big job on, you can just do it, and then not use it for the rest of the month if you want, that sort of thing.

Katie (04:48)

Yeah. And do you prefer that, rather than a daily limit?

Noel (04:55)

Yeah, I think the five hourly is a bit strict, because I can get to lunchtime and it'd be like, right, you can't use it now until two o'clock. That's a bit annoying. But with Anthropic, though, they did give me some free credits a while ago now. I don't know why — I kept having little buttons popping up saying claim this $90 offer, and I was like, okay, I'll claim the offer. So I had loads of extra usage. So you can pay more to use it if you want to, without upgrading your plan — you can just add money in.

Noel (05:24)

But yeah, Fable 5 chews through that money in no time. So I hit my limit and then got through $50 of their claimed credit before it came back on. That's what I mean — it just doesn't seem that you get so much usage.

Katie (05:48)

Okay. That's really interesting. Okay, any other updates that we should be aware of?

Noel (05:52)

I don't believe so this week. Obviously we're going to go through some more in a bit more detail now, but other than that, I think everything was kind of quiet. But yeah, this next bit — we did ruffle a few feathers, I think.

Katie (06:08)

Okay. So yeah, for this week's podcast episode, there have been new AI LLMs released. So Noel, do you want to tell us a bit about them?

Noel (06:22)

Yes. So I guess just to go back a little bit — last year we did an episode about DeepSeek. I think it was episode four or five of our podcast that we'd ever done. And I remember at the time, it kind of shook the market a little bit. This DeepSeek model was cheaper, and it was probably a bit more capable than the frontier models from OpenAI, Google and Anthropic. So it caught everyone by surprise, and we've had that again this week from two other Chinese labs. We've had the release of Kimi K3 from Moonshot AI, and Alibaba have just released Qwen 3.8. And they're very, very capable models — a lot more capable than maybe these bigger companies thought they would be. So yeah, we're hitting that again.

Katie (07:24)

Okay. So Kimi K's obviously been going for a little while, hasn't it? And I think we've talked about Kimi K on the podcast before, but this is now Kimi K3.

Noel (07:34)

Yes, yeah. I think it was 2.5, I think, the first time we mentioned it. But that was last year.

Katie (07:42)

Okay. Yeah, so do you want to tell us a bit about Kimi K3 — what it does, why we'd want to use it, or what we would use it for, things like that? Because I think it's always really helpful. When there becomes so much choice, it's kind of overwhelming almost, and it's like, which one is going to be the best for me? What are my options?

Katie (08:13)

So many people, I think, just stick with ChatGPT because it's something that everyone knows. It's, I guess, trusted. But the problem with ChatGPT is, if you're using it and everyone else is using it, everything looks and sounds very similar. Like, you always know when a graphic or a poster has been done by ChatGPT, because they all look exactly the same.

Noel (08:46)

Yes, yeah, definitely. So with Kimi K3, it's very, very capable — it's almost at Fable 5 level intelligence. This has kind of caught people off guard. So at the moment, with OpenAI and Anthropic, you've got Fable 5 at the top end, and then 5.6 Sol was just as good, but just a bit below. But what Kimi K has done is they've slid in the middle somehow. They've managed to sneak in, and everyone's like, what? How have they managed this? Surely this can't be right. But all the benchmarks are all independently verified and that sort of stuff. So yes, it is almost as powerful as Fable 5.

So it's good for creating apps or websites, it's good at creating visuals, and it's also good at being the brain in an AI agent, things like that. It's really, really good at that, because I put it through Clyde Bench on the Clyde platform, just so my users understand which models are the best, and it aced it. It got a hundred out of a hundred score. So yeah, it was very, very good at those sorts of things. I would say it's incredibly powerful. It's quite quick as well — like I say, Fable 5 is very slow, and 5.6 Sol is also quite slow. So there are a lot of benefits, potentially, to switching to Kimi K3.

Katie (10:21)

And how does it compare price-wise?

Noel (10:23)

So this, again, is where it shook everyone up. So you can sign up to their platform, like you can with ChatGPT and Anthropic and that sort of stuff. You can do that, but most people are going to use it on the API, with agents and coding and things like that. In terms of API pricing, it sits at the same level as Anthropic's Sonnet — it's the exact same price.

Noel (10:53)

So it's almost as powerful as Fable, but it's two tiers lower in terms of pricing. So a lot of people, especially big businesses that have moved to Fable 5, are now looking at this and going, well, why are we spending all of this money on Fable 5 when we could have just as good for a third of the price? So that's the interesting way people are looking at this now. Is it worth the risk to switch everything over to Kimi K? Because it is open source as well — you can host it and edit it if you wish, if you had that sort of computing power. So there are lots of people looking at this and thinking, is it worth me sticking with OpenAI and Anthropic? Should I look somewhere else?

Katie (11:43)

So we've got lots of pros towards Kimi K3. What would be some cons towards using Kimi K3?

Noel (11:48)

So it's similar cons to when we went through DeepSeek. Moonshot AI is a Chinese business corporation based in Beijing, I believe, so they have to comply with Chinese cybersecurity laws and things like that. So there is a potential risk that if the Chinese government see Moonshot as doing stuff that's maybe a bit sketchy, or they're not quite happy, they can demand to see all of the data. That could be stuff that you've put in and typed in on their chat interface, it could be on the API — they can demand access to all of that data and go through it all, and potentially see personal things about you. If you put things in about your clients, then that could also be seen by a foreign government, which for most businesses is probably not ideal.

Especially mine — I think I'd have some customers that would be very angry if that happened. So there's a bit of a risk. Chances of it happening, I don't know, you'd have to weigh that up. But they can't just go in and go, I want to see everything. They have to at least have some sort of suspicion that something's going on on the platform that shouldn't be allowed.

Katie (13:06)

Okay. Because I think everyone is a bit frightened of the Chinese cybersecurity rules and regulations. But how does that compare to companies who are based in, say, America?

Noel (13:28)

Yeah, so it's almost very similar, in a way. They can demand data and things like that. I think there have already been court rulings before that I've read about, where they said, well, we want to see the ChatGPT history and stuff like that, to see what they've been looking up and checking. And then demanding that data — whether or not it was given up, I don't know, or what they got was kind of brushed under the carpet a bit; they didn't really say what they gave up. But I guess it's the same on any sort of data platform — any government or law enforcement can go, I need to see what they've been up to, and whether or not they give it is another thing. So yeah, there are risks everywhere.

Katie (14:49)

Yeah. Because I think we're very used to the idea that your search browser history can be looked into and can be used against you as evidence, or whatever, if they're trying to build a case for something — I think we're all aware of that. And because obviously AI and LLMs are so new, we're not quite aware of that — we've not got it in our heads. A bit like, I guess, when we first started using Google — we would just type in any old thing and not really think about it. Whereas now, when you have those unhinged conversations with your friends and they're like, look that up, they go, well, I'm not having that on my search history.

Noel (15:17)

What do you talk about with your friends? God, they're dodgy.

Katie (15:25)

Nothing dodgy going on! Nothing dodgy going on. But yeah, I think we need to remember that there is this online digital footprint on whatever we use, isn't there?

Noel (15:28)

There is, yeah. And when it comes to businesses, you've got to weigh up the risks, haven't you? And if it's worth it, then do it. What you could be wanting to do with Kimi K might not involve any personal data, and you can probably just go, yeah, that'll be fine, we can use it for that particular use case. But am I going to let it have my entire company access, on my CRM, via an agent using Kimi K? Maybe not.

Katie (16:13)

So you wouldn't use Kimi K to run an agent?

Noel (16:18)

It depends on the agent usage and what it has access to. If it's like a research agent, fine, that's easy enough. But if it was going off and creating emails to send to customers, or reading emails, or reading a CRM or databases which have personal or potentially classified information, then no, I wouldn't. I'd also be a bit worried about the mainstream ones as well with some of that sort of stuff.

But like I say, these models are open source. So if you did have the computing power — and I have asked Claude, could I run it on my PC? And he went, don't be stupid, it's far too big for that. This model is massive compared to others.

Noel (17:12)

But you could, if you had the capital to do so — you could host it on the cloud, and then whatever goes in and comes out, all your stuff is ring fenced. At that point, it's fine.

Katie (17:16)

Okay. But you've got to have a pretty big server.

Noel (17:28)

Very big, yes. You need deep pockets, unfortunately. But it is getting better, though. I did nerd out and read through a bit of a paper that's coming out, and they're talking about how to make AI models more efficient going forward. Because obviously, at the minute, the more powerful they get, the more computing resource they're going to need. Whereas they're kind of looking at it and going, well, we want it more powerful, but less computing power. So hopefully we're going to get to that point where it's going to start creeping down, and we'll be able to have Kimi K3 on our mobile phones. One day. It'll probably be Kimi K10 or something.

Katie (18:09)

Yeah. Okay. So what's the other update that you wanted to talk about today, Noel?

Noel (18:19)

So the other update is Alibaba. Obviously another Chinese AI lab, and they've brought out Qwen 3.8. And again, this is ruffling a few more feathers, because they've squeezed into that GPT 5.6 and Fable 5 gap — they've snuck in with Kimi K.

There's no talk about pricing yet with that one just yet. It only got released on their web app, I think it was two or three days ago, so I don't think the API is out just yet, but I am keeping an eye out for it. Looking at their past models, it's probably going to be similarly priced to Kimi K — it's similar sorts of capabilities. So again, it comes with the same sorts of risks.

But what I would say with these two models in particular is, if you're looking at creating websites, or you want something that's a bit more visual and a bit more exciting to look at — and not the usual AI rubbish that comes out — then these models are really good at creating that sort of stuff. They're very, very good at creating even 3D websites and stuff like that. So if you want to go space age on your homepage, then these models can definitely do that. And they do far better than Fable 5 in those sorts of areas. So yeah, really, really interesting.

Katie (19:46)

And how does that compare, for the security and data, to Kimi K? Is it along the same lines?

Noel (19:55)

It's exactly the same. Yeah, it's the same risk. They are a Chinese corporation, so they do have to comply with the cybersecurity rules from China. But with these as well, they also have data centres around the world. I was looking it up earlier, and Kimi K have their own servers in Frankfurt, so they can look into more like GDPR and EU regulations, so they can have that sort of data sovereignty within Europe or the US and around the world.

Noel (20:25)

But yeah, the same risks do still apply. Even if your data is in Frankfurt, the Chinese government can still request access to it if they wish to.

Katie (20:42)

Yeah. Okay. Anything else you want to add?

Noel (20:51)

So I think that's really all of the big news. And yeah, I'm kind of excited to try them out. I've seen the visuals and stuff that they can produce for websites, so I'm kind of looking forward to giving that a go and seeing what I can do.

Katie (21:07)

Yeah. Will you put some of that in the free LinkedIn group, AI Automations for Business, once you've had a little play around with it, and let people see what those graphics and images look like?

Noel (21:22)

Yeah, yeah, definitely. Because I was getting angry with Fable 5 the other week, actually. I was just like, this just looks rubbish — what are you doing? Come on, let's think outside the box. I've described something and you've gone and just given me the generic rubbish. So yeah, I'll try the same prompts and put them in there with Kimi K3, and when I can get access to Qwen 3.8, I'll do it in there as well.

Katie (21:33)

So if you're not in the free LinkedIn group and you do want to see what Noel's creating, then please come and join. Everyone is completely welcome. You can come and let us know what you've been using AI or automations for — come in and showcase what you've been doing. It's very, very welcome, and we always love seeing how people are using AI and automations in their business.

What I wanted to say, though, was: have you noticed, Noel — or for anyone who is listening, with Anthropic, with Claude — since Fable 5 has been released, has anyone noticed that Sonnet 5 isn't as good?

Noel (22:33)

You know what, I did use Sonnet 5 yesterday, just to help me with my usage, and I would say that it was a stark difference in terms of use. It just made a mess of what I wanted it to do. But I find, actually, I tend to default back to Opus 4.8 a lot, and I'm finding that's also maybe not as smart.

Noel (23:17)

I don't know if it's just that my brain is expecting Fable 5 outputs and then getting not so good outputs. So yeah.

Katie (23:23)

Yeah. Because, you know, we had the conversation — I think it was earlier this year — where both of us had just had enough of ChatGPT and we decided to move to Claude. Was that this year or was it last year? I can't remember.

Noel (23:37)

I think it was the very start of this year.

Katie (23:47)

Okay. I feel like some of the outputs that I'm getting from Claude Sonnet 5 right now are like those I was getting from ChatGPT earlier this year, where it's not listening, it's not remembering things, or it's just not as good as it used to be.

Noel (24:06)

That's interesting. Yeah, because it's kind of funny — even with Opus 4.8, the written content was probably not exactly what I wanted. It used to be really good. When it first came out, I was like, wow, okay, this is brilliant. Whereas now, I'm kind of questioning it a bit more than I used to. So yeah, maybe they are getting dumber over time. Who knows?

Katie (24:15)

I don't know. But it'd be really interesting to know other people's opinions as well. So yeah, come and let us know in the free LinkedIn group, or of course you can always email us, hello@makeautomations.ai. Drop us a line, let us know your thoughts — we're always really interested to hear your opinions as well. But thank you so, so much for listening to this week's episode. We hope you've enjoyed it, we hope you've learnt something new.

Katie (25:07)

And we will catch you next time for another episode very soon.