How Beats Stronger AI Without Using It?
39sThe paradoxical claim that an AI can beat a stronger model it doesn't even use sparks curiosity and debate.
▶ Play Clip"Title promises destruction but delivers a balanced overview with a heavy sales pitch for the creator's own system."
Sakana's Fugu Ultra V1.1 is a multi-agent AI system that claims to outperform Claude Fable 5 on coding and reasoning tasks without even using that model. The video explains how a team of specialized AI models can surpass a single powerful model through collective intelligence, while also discussing trade-offs like speed and cost.
A Tokyo-based company called Sakana released Fugu Ultra V1.1, claiming it beats Claude Fable 5 in hard coding and thinking tasks.
Founder David Ha (Hard Maru) posted that the new version improved by 7.9 points and beats Fable 5 without including it in the model pool.
Unlike typical single-model AI, Fugu uses multiple models working together: one thinks, one acts, one checks, producing a final result.
Added the newest frontier models to the pool, boosting scores by up to 7.9 points on benchmarks like Program Bench and Terminal Bench.
Despite the improvements, Sakana did not raise the price for the better version.
The core insight: a team of good AIs working together can outperform one great AI, even if that great AI isn't in the team.
Sakana calls this 'collective intelligence' – many AIs collaborating like a school of fish.
Users report Fugu is slow and expensive to run. About 75% of reactions were positive, 25% negative.
The benchmark results come from Sakana themselves, so they should be taken with caution until confirmed by third parties.
The real lesson: the ability to quickly integrate new tools into one system is more valuable than any single AI model.
Fugu Ultra V1.1 demonstrates that multi-agent systems can rival top single models, but trade-offs in speed and cost remain. The key takeaway is to build a flexible system that can leverage any new AI tool rather than relying on one model.
What is Fugu Ultra?
A multi-agent AI system that uses a team of models to complete tasks.
01:22
What does 'Sakana' mean?
Fish in Japanese.
01:51
How much did scores improve in v1.1?
Up to 7.9 points.
02:22
What are the main trade-offs of Fugu?
Slower and more expensive than single-model AI.
03:29
What is the key principle about AI tools emphasized in the video?
The system around the tool is the moat, not the tool itself.
06:25
Who is the founder of Sakana?
David Ha (known as Hard Maru online).
00:17
Teamwork beats one genius
Core insight that multi-agent collaboration can outperform a single powerful model.
02:51Collective intelligence
Introduces the concept of many AIs working together like a school of fish.
03:17System over tool
Key strategic advice: build a flexible system rather than chasing individual AI tools.
06:25A team of good AIs beats one great AI
Summarizes the main takeaway of the video.
07:08[00:03] really beat Fable 5? Let's find out. Here's the wild part. A company in Tokyo called Sakana just shipped a new AI. It's called Fugu Ultra V1.1, and they say it now beats Claude Fable 5 in hard coding and hard thinking tasks. But
[00:17] here's the twist. It beats Fable 5 without even using Fable 5. Read those words from the founder himself. His name is David Ha. He goes by Hard Maru online. He posted this on July 23rd. He said his team pushed the score up by 7.9
[00:30] points, and they beat Fable 5 without Fable 5 in the pool. That's not me guessing. That's the guy who built it saying it out loud. So, how does one AI beat a stronger AI it isn't even using? That's the question. Stick with me
[00:43] because the answer changes how you should think about AI tools. And I'll tell you now, I'm going to run this thing inside my Agent OS. My Agent OS is the system I built to run every AI agent I use in one place. One screen, one
[00:56] shared brain. So, when a new tool like Fugu Ultra drops, I don't start over. I just plug it in. More on that in a sec. First, let me explain what Fugu Ultra even is. Hey, if we haven't met already, I'm the digital avatar of Julian Goldie,
[01:09] CEO of SEO agency Goldie Agency. Whilst he's helping clients get more leads and customers, I'm here to help you get the latest AI updates. Julian Goldie reads every comment, so make sure you comment below. And I'll keep it simple. Most AI
[01:22] you use is one brain. You ask, it answers. One model doing the work alone. Fugu is different. Fugu is not one brain. It's a team. When you send it a job, it picks the best AI models for that job. It hands out roles. One model
[01:37] thinks, one model does the work, one model checks the work, then it hands you boss who hires the right workers for each task. You don't meet the workers, you just get the finished result. That's Fugu. And here's a fun fact. Sakana
[01:51] means fish in Japanese. The idea comes from a school of fish. Lots of a fish moving like one big animal. That's the whole idea behind Fugu, many AIs moving you? Because you don't have to pick the best AI anymore. Fugu picks for you task
[02:07] by task. And when a better model comes out, Fugu just uses it. You don't lift a finger. Okay? So, what's actually new in V1.1? Sakana said they added the newest frontier models to the pool. That one change pushed scores up across every
[02:22] test they showed, up to 7.9 points higher than the old version. The biggest called Program Bench, one called Terminal Bench 2.1. Those tests check if an AI can really write code and run real tasks in a terminal, not just talk about
[02:38] it. And the price, same as before. They didn't charge more for the better version. That's rare. Usually better means pricier, not here. Now, here's the headline claim, beating Fable 5 without Fable 5. Let me make that make sense.
[02:51] Fable 5 is one of the strongest models out there. But Fugu doesn't have it in the team. So, how does the team beat the star player who isn't even playing? Simple, teamwork beats one genius on the hard stuff. A few good models working
[03:04] together, checking each other, that can top one great model working alone. One AI writes, another spots the mistake, another fixes it. The answer that comes out is better than any one of them alone. That's the big idea Sakana keeps
[03:17] pushing. They call it collective intelligence. Fancy words, plain meaning a team of good AIs can beat one great AI. Now, let me be straight with you because I'm not here to hype. Not everyone loves it. Some users say it's
[03:29] slow. One guy literally posted that it thinks too slow. And he's got a point. When you run a team of AIs, it takes longer than asking just one. There's more back and forth. Some users say it's pricey to run. Running a team costs more
[03:41] than running one. That's just true. About 75% of the reactions were positive, 25% were not. So, it's good, but it's not magic. Slower and costs jump in. And one more honest note, these scores come from Sakana themselves, not
[03:56] from a neutral judge yet. So, I take the numbers with a grain of salt until other pause. Let me help you actually use this stuff. If you want to plug tools like Fugu Ultra V 1.1 into your work the right way, come see how we do it inside
[04:10] the AI Profit Boardroom. We've got a lot of members already running multi-agent setups just like this using one system to route the best AI to each job. Inside, you get four coaching calls every week where you can ask live, "How
[04:23] do I wire Fugu Ultra into my workflow?" The Agent OS is what you want, and it ships as a complete zip file. Download it, install it. You get daily step-by-step tutorials showing exactly how to set up tools like this to save
[04:36] hours a week. And you get a 30-day roadmap, so you're not guessing what to description, or go to the AI Profit Boardroom. Okay, back to it. So, who is Fugu Ultra actually for? Let me make it real. It's built for hard jobs, not
[04:50] "Write me a quick email." I mean, big, messy, multi-step work. Sakana says it's checking science papers, and looking at security problems. Stuff where getting it right matters more than getting it fast. Here's a simple way to picture it.
[05:04] Say you're building a big landing page for a business, one that has to explain a lot and convert visitors into customers. A one-brain AI might rush it. Fugu sends it to a team. One drafts, one reviews, one polishes. You get a cleaner
[05:17] result. That's the kind of job it shines on. And here's how it fits your day. You tools. You want one thing that just picks the best one for you. That's the promise here. Less choosing, more doing. Now, let me show you how I'm using it.
[05:30] This is the part that matters. I'm running Fugu Ultra V 1.1 inside my Agent OS. Remember, my Agent OS is my one control room for every AI agent I use. Instead of 10 open tabs, 10 logins, 10 tools I forget about, it's all in one
[05:46] place sharing one memory. So, when Fugu Ultra dropped, I didn't rebuild anything. I just added it as one more engine. It sits right next to my other agents. It uses the same shared brain, and I can hand it a job from the same
[05:59] screen I use for everything else. Why does that matter? Because tools change every single week. A new best AI drops constantly. If you build your whole when the next one comes. But, if you build one system that any tool plugs
[06:12] into, you're never stuck. New tool drops, plug it in, move on. That's the real lesson under all of this. The tool is not the moat, the system around the tool is. Fugu Ultra proves it. It's not one model, it's a system that uses
[06:25] whatever model is best right now. My Agent OS works the same way. One system, me give you one more real example of how I'd use this. Say I want to write a deep, detailed guide for my community. I'd hand that job to Fugu Ultra inside
[06:41] the Agent OS. It runs the team. One writes, one checks facts, one tightens I go do something else. That's the whole point. The work gets done without me babysitting every step. And look, the speed thing inside a system like this,
[06:54] slow matters less because I'm not sitting there watching it. I hand it the background. I come back to a finished result. Slow only hurts when you're staring at the screen waiting. So, let's bring it home. Fugu Ultra V1.1 is real.
[07:08] It's a team of AIs sold as one. The new version jumped up to 7.9 points, mostly in coding. It claims to beat Fable 5 without even using Fable 5. It's slower and it costs more. That's the honest trade. And the scores still need outside
[07:22] testing before we treat them as gospel. But, the big idea is the thing to hold on to. A team of good AIs can beat one great AI. And the smart move isn't to chase every new tool, it's to build one system that any new tool drops into.
[07:35] That's exactly why I run mine the way I do. Here's my last ask, and I'll keep it tight. The whole reason I can plug in a brand new tool like Fugu Ultra in minutes is my Agent OS, and you can get the full Agent OS inside the AI Profit
[07:49] Boardroom, the complete zip file ready to install, not a demo, the real thing. You download it, you set it up, and you've got the same control room I use to run every AI agent from one screen. Plus, four coaching calls a week to help
[08:01] you wire in new tools like Fugu Ultra as they drop, daily tutorials, and a 30-day roadmap so you know exactly what to do. Come grab the Agent OS zip file inside the AI Profit Boardroom, link in the comments and description. And if you
[08:16] want the full process, the SOPs, and 100-plus AI use cases like this one, join the AI Success Lab. It's our free AI community, links in the comments and notes from there, plus access to our community of 85,000 members who are
[08:31] crushing it with AI. That's Fugu Ultra V 1.1. A team of eight is beating a giant it didn't even use. Now you know how it works and how to use it.
⚡ Saved you 0h 08m reading this? Transcribe any YouTube video for free — no signup needed.