Fugu: NEW Japanese AI DESTROYS Fable 5?

Japan are absolutely dominating with a Fable 5 level model it's called Fugu Ultra and it can match the performance of Fable Mythos literally just got dropped a few hours ago today and I've already tested it out a bunch I'm going to show you what I learned what we've gained from it and how it scores on the benchmarks with GoldieBench so we're going to test this out and if you've never used this before you're not sure what Fugu is what it is etc I'll show you some stuff that built with it and then in a second we can run through the benchmarks and everything else so if we have a look here you know this is all one-shot stuff so the way that Fugu works is basically it has like a bunch of AI models that work together some open some close and then together they can actually create answers in parallel and what happens after that is a judge actually fuses the answers together and you can see like the quality of this stuff it's pretty nice like this website pretty good and we can see another example here look at this little galaxy right there little live universe lots of number one it looks beautiful but number two it's actually really cool like it looks amazing visually as well like it's fully animated it's just spinning round as a live spiral galaxy and then we also have this example of an inner solar system now what this is essentially is something called Fugu Ultra and it's a multi-agent panel API that competes head-on with OpenRouter Fusion so OpenRouter Fusion just released their own version of this as well last week and both of them work with a very similar system so what happens is you ask something you get one prompt then it's going to give it to a bunch of models so for example like GPC 5.5, GLM 5.2 or Kimi K207 at the same time from there they synthesize the votes and you get one single answer so basically what the students like having a bunch of really smart brains AI brains to work together answer something and then you get one answer back now there's two different models here so there's Fugu Ultra and there's a standard version of Fugu as well so this is a multi-agent system that works together to get better answers now the way this works you can get one API to access them all so you can coordinate them all together we actually built that system into our agent OS so if we go to our agent OS over here you can see that we've added Sakana Fugu and so this is a system where we can have all of our builds in one place we can access it whenever we need to we can have a look at what we've previously created with this stuff and then also the great thing about this is that we've got this ready to go so if we need to ask a board of AI let's say for example we're creating something awesome we've got it ready to go inside the agentic OS and also have Fusion as well so these are two different models working together in the same sort of way and Fusion is pretty cool as well if you've never seen it before it can build some awesome stuff as you can see right here I mean it's created for example City Fusion which is a live city simulation with people and cars and all that sort of cool stuff and you can see some other stuff that will build like cool games and that sort of thing so they're both really cool models you've got Fugu which is fast so that's designed to have performance and latency and then you have Fugu Ultra which takes absolutely ages to respond but it's optimized for performance right it's designed for like better performance now if you look at the benchmarks here you can see the Sakana is outperforming Fable on many benchmarks like for example Terminal Bench 2.1 Reasoning Tests, GPQA Diamond, Live Codebench so on a lot of the benchmarks Fable 5 is actually being performed by this tiny little AI lab now what are the strengths of this well SWE Bench Pro GPQA is published frontier level benchmarks it's also vendor agnostic so you can opt out of specific providers when you use this API and it's also OpenAI compatible so you can drop it into your existing tools like we did with the AgentOS the trade-offs are that because it's a panel panel orchestration that makes it slower and also it's newer than Fusion so there's less information about this pretty hard to find information I mean it just dropped today for example Also we've tested it with some stuff and it actually got truncated on the response so what that means for example so we have a look at this Voxel Temple Run game that we tried to build it doesn't work and that's because the answer actually got truncated it hit a 16k max token ceiling when it was being spawned which means that the game actually doesn't work so it's just something to be aware of so you can wear we I literally waited about 25 minutes to generate that response and it didn't even work now if you're wondering okay how does it perform against Opus 4.8 for example it pretty much crushed Opus 4.8 on every single thing that we created and you can see all those tests over at GoldieBench the only thing it didn't outperform at was the Voxel Run game because it just literally didn't work and also the API is quite limited so for example we after four tests we actually hit the limit and couldn't generate anything else for like five hours so even on like a $200 a month set plan we couldn't keep using the subscription because we ran out of tokens literally said you can't generate any more so that's something to be aware of as well now if you want to indicate how does it perform on the leaderboard bear in mind we still got to do more tests so it's going to take a bit longer to get the full explanation or keep testing it but it's actually scoring number six right now I think once it's generated more outputs and we've scored more tasks it'll probably rise up it's just because that Voxel Run game just didn't work at all and that's what brought it scored down. Fusion is still at the top that's created the best outputs for us so far it's created some pretty amazing stuff if you want to see some examples of this let's check this out I mean all sorts of amazing stuff here like a cool Raycaster game as you can see right here a lot of fun to play this cool Dragon Realm game as well which is pretty amazing I mean look at this the quality of this just absolutely amazing and here's the way that I look at it it's like you know we can't access Fable 5 but we can access Fable 5 level outputs however it's nowhere near as smooth as using Claude Fable 5 after testing it a bit more the problem with all these models is like they take so long to reply and also it's one shot which means for example like if you're creating outputs with Fugu you're gonna have to wait like 5 to 10 minutes and then you can't really go back and forth with the CLI whereas for example when we were using Fable 5 you could quickly say okay this is not good enough this is good enough and you can go back and forth with it in a much more fluid way right you're not waiting 5-10 minutes for each response so Fable 5 was still absolutely amazing not so much in the stuff that it created which was already good but really in the speed of it and I think that really counts when you're building stuff out like if we're trying to create cool stuff inside the agent OS we're trying to build new features etc of course like the Claude CLI is going to be way faster and easier to build stuff whereas for example you might build something with Fusion or with Fugu and you might be waiting like 10 or 15 minutes and then it doesn't even work so just something to bear in mind there that's for example why a lot of the tests we actually tested before don't work yet or haven't been created yet so something to bear in mind right there one big difference between OpenRooter Fusion because you might be saying okay well should I use Fusion or should I use Sakana so Sakana has a flat rate subscription and you can choose which subscription you want and that's quite a cool way to use it whereas with for example OpenRooter Fusion is pay-as-you-go only you're paying per usage however what that also means is that if you're paying per usage you can keep using it you can keep building with it whereas for example you're going to hit the limits very quickly with Sakana from what I've seen and then you're sort of frozen out for the next five hours so it's up to you and do you want cheap or do you want fast if you want fast probably go with Fusion if you want cheap then I would go with Fugu now if we actually have a look on OpenRooter you can't find Fugu so the way that I got access to it was through Sakana.ai I also saw that some people were from the well based in Europe and they actually couldn't access Sakana they couldn't use it I think that's due to like GDPA or something like that like privacy rules and restrictions so that's something to bear in mind as well as like you can get access to OpenRooter Fusion pretty quickly but it doesn't seem like Sakana is available everywhere this has also happened with Fusion so for example if we click on this game here it doesn't work at all and it's the same sort of problem which is sometimes the outputs get truncated and then you don't get the full code or the full answer you can't really use it so something to bear in mind there as well like if you want just the smoothest experience but but maybe not the best of the best of the best outputs then you go with Opus 4.8 and bear in mind like when you're using Opus 4.8 the great thing about using that is that if something is not good first time around it doesn't really matter because you can go back and forth with it as much as you want and it's super fast to go with right and that is going to use way less tokens than for example if something breaks on Fusion well Fusion or Sakana you can't really go back and forth in real time now I think the best way to use all of this is inside an agentic operating system like you can see right here and this is just an amazing way to plug them all in so what you can do is you can like plug in a prompt into Sakana then you can come back later and you could also do the same with Fusion now whilst you're doing that you could also speak to Cloud CLI and the great thing about that is you can have multiple agents working in parallel and you don't get blocked there's no constraint holding you back in terms of time and tokens you can build multiple things out at the same time in parallel we've even got for example like Paperclip and the AI Agent Mastermind and the Pipeline right here ready to go so overall Sakana pretty cool idea I think that's the way everything's moving if you look at Fusion if you look at Sakana they all work pretty nicely together at the same time I would be prepared for slower outputs and sometimes the answers are going to get truncated which means the outputs are not as good as just using Cloud directly so the best way to use this is to have multiple models working in parallel and then if something breaks no problem you got something else that's working in the background you don't get held up so if you want to get my full agent operating system with Fusion with Sakana built in with the Cloud CLI as well you can get that inside the AI Profit Boarding link in the comments description or just go to the aiprofitboarding.com inside the community you can ask questions get help and support I personally answer these questions every day with video tutorial inside the classroom you can get access to the Agent OS system and you can also see our full guide on how to use Sakana and Fusion with a step-by-step guide and a video tutorial we've also got a video tutorial and a new update for Agent OS and you'll get the zip file for that too and then every time something useful comes out we share it also if you're a complete beginner you can go through the beginner to expert course right here if you want to jump on weekly coaching calls jump on a zoom call ask questions share your screen meet the community bounce some ideas of people you can get that right here and inside the map you can actually meet people in your local area who are building with stuff like Fusion, Sakana and everything else inside the map right there so feel free to get that link in the comments description or go to the aiprofitboarding.com thanks for watching bye bye