What is Sakana Fugu? Overview & quickstart

0:00Today, I want to talk a little bit about Sakana Fugu, and we're going to discuss what Sakana Fugu is, why it's pretty cool, and then I'm going to show you how to run it. Two very tactical ways to actually use this thing. Actually, three ways. We'll add a bonus way just because I'm feeling generous today. So, let's

0:14let's jump into it. What is Sakana Fugu? Bold claims on the lander here. One model to command them all. Um and I think it's a little confusing just to understand what Fugu actually is. Uh because if you look at their description here, and this chart is maybe a good explainer. Um it claims that Fugu kind of orchestrates closed and open models, or leverages closed and open models. What it means to orchestrate a model, I think, is kind of up for debate here. But that Fugu itself can recursively call itself, right? So, like that's a little confusing off the bat. Let's break it down. There's a research paper that the

0:54model's based on. Um The authors kind of share this, right? And this is the general architecture. I think this is already a little bit more clear. So, there's a user query, somebody sends through a prompt. Um there uh is some like processing that happens here. This concat step SLM, I don't really know what SLM is. Um it doesn't seem integral to what's going on. And then uh there's a lightweight model that's kind of like a router. And so, what this router does is it takes a pool of LLMs and it assigns

1:24them roles. Um and that role might be thinker, worker, or verifier. Um From there, we're passed to the actual verifier. Uh and there's a bit of a loop here, right? Where the verifier can do stuff. Now, this is maybe a little confusing, so I broke it down even more. So, basically, we have a query that goes to a coordinator. That coordinator sends in um uh the message and assigns an LLM to these roles. So, sometimes Fugu's like, "Hey, you know, like we don't actually need to do this. We're just going to solve it and just run through it. It's a very simple example. Other times it's

2:01going to assemble like a team of experts. And when I say team of experts, it just means it's going to act as an agent orchestrator. It's going to take it's going to be an orchestration model who then assigns tasks, maybe thinking, doing the work, verifying the outputs to these different models. And Sakana says, "Hey, we use a combination of open and closed source frontier models to do this work." So, like it might be Codex, I be GPT-5. It might be Opus-48. It might be, I don't know, GLM-52. They don't explicitly say that, but we can assume they're using some open source models as well. And most

2:35importantly, Sakana mentions that it can use itself. So, it can recursively kind of call this loop, if you will. So, this could actually just be like one subset of a larger agent run. That's like a little confusing and a little meta, but that's how it works. And from there, we ultimately derive a response that we're going to return to

2:55the user. So, at the outset, you might be saying, "Okay, is this thing like Claude Coder? Is it like Opus? Is this thing like Codex? Or is it like 5.5?" The answer is yes. It's an agent, but we're calling it through an API, right? So, this is being called from an API. We don't install Sakana Fugu. We call Sakana Fugu. We call it as an API and we use it with another harness. So, you can use Sakana Fugu Ultra or Fugu. So, Ultra is just like the heavier version that does more stuff. You can use these models with like a open code. You could actually use them with Codex, which

3:29we'll talk about today. It's on the docs. So, next, pricing. What does pricing look like for this model? Um Cost is five um dollars per million input tokens, $30 per million output with 50 cents per million cached input. Identical pricing to GPT-5, um a little bit more expensive than a Opus-48. Now, the important thing here, right, is that they're claiming on the website that um Fugu is going to like do stuff that rivals Fable, right? So, the reason this is a big deal is because we're saying, "Hey, you know, Sweet Bench Pro um Fugu

4:07Ultra is outperforming these models. Fugu Ultra is outperforming Mythos Preview. It's outperforming Fable 5 on Live Code Bench and a couple of these other things." These are lesser-known benchmarks. Uh I think Terminal Bench is an important one, but again, we're we're saying, "Hey, better results than Fable on Terminal Bench." Um maybe not on Sweet Bench Pro. So, the claim here, right, is that we have GBT 55 pricing with potentially Fable 5

4:34performance in some areas. So, that's like that's like the claim, right? How do we actually use this thing? Cuz I think this is this is confusing. So, there are a couple ways to use this. One way is through the Sakana API console, right? So, you go to Sakana AI and you create a subscription or you create an API key and you pay them and you're going to use the model through their uh console directly. Another way is through OpenRouter. So, OpenRouter has access to Fugu Ultra. So, you could um create an OpenRouter API key um in your profile and then use that OpenRouter API to call Fugu through

5:09something like an OpenCode. I'll show you what that looks like. So, if you wanted to use OpenCode, actually, it's the easiest. Um I use Conductor. I work at Conductor. Uh so, typically, what that would look like is we go into our settings, we'd go to our harnesses, and then OpenCode. And we configure a provider. So, maybe we go into OpenRouter, we drop in our API key, we save that, and then we can configure a model. Since we have OpenRouter configured, we could go Fugu

5:38Ultra. Selecting Fugu Ultra is going to allow us to then select it from the model picker. That looks different if you're using OpenCode, but that's the general premise, right? You bring your own API key with a provider that has access to Fugu, and now you have this thing. Now, the other thing you could do, rather than running it through open

5:55code, is you could run it through Codex. What does that look like? Um Sakana actually has a nice little getting started guide and explains how all of this stuff works. It looks a little dense and a little verbose, but it's actually simpler than than it first seems. So, what's happening here is that Fugu's going to override a bunch of your Codex setups. It's actually not overriding anything. It's just accenting or

6:18augmenting your existing Codex setup. So, assume you already have Codex installed. If you don't, install Codex the CLI on your computer. And what you're going to do here is you're going to run this curl command, this like standalone installer. Um they might actually have this, yeah, right here. So, if we open like our terminal, um and we uh I'll see the up one directory, if we run like this, something stuff is going to happen, right? And you can see what's happening

6:42if we read through this. Um is it's overriding values in config.toml for Codex. Um and it's creating some uh different paths. Um now, the thing that's confusing here and that I actually bumped into some of the issues is that it's unclear to me exactly where to set environment um keys. So, what you'll want to do is open up uh something like a a file editor, maybe like Z, and what you'll see now is that in Z, we have um in our codex.codex, this is at the root directory on your computer, config.toml, we have a model

7:18providers Sakana. And what that's saying is we're adding a new model provider with the Sakana API, and we're looking for the environment variable Sakana API key. Now, one very way to make very easy way, rather, to make sure you get this environment variable is to create an API key in your Sakana dashboard, and then drop it in a .env, which I have here in the in the same directory. So, now Now storing our Sakana API key in the root codex directory. And when we run codex on our machine and select the Sakana

7:47model, it'll get picked up. So, this is kind of the second way to run Sakana and it's through codex. So, assuming you ran that bash command and pulled that down, you're going to get this like Fugu JSON file which has more Fugu config, a Fugu config file. Um these are all just telling the model

8:04how to behave. What this enables, right? Assuming you have that API configured and you can do this by running something like echo uh Sakana API key. Um don't do that cuz it'll print your API key. Like don't do it while anybody's watching. Uh but you shouldn't need to do that and it actually might not be set because it's going to be set by codex. So, if I run codex P Fugu now, that's going to say, "Hey, I want to run

8:27codex with the model Fugu." And I can say, "Who are you?" And we should get, "I am Sakana Fugu." Right? Ideally. I'm Fugu. Great. So, um that's the second way to run it, which is to run that bash script, configure Fugu, and then fire up a codex session. Now, I said three ways. I promised three ways, right? And the third way is to run curl, this guy. So, as I mentioned, it's a model. We don't actually need to call the model um with a harness. We can just call the model directly, right? And this is the same way you'd put Claude in your web app or

9:05you'd put chat GPT in your web app to like do stuff. We can um actually just make a curl request to the Sakana endpoint. And so, I actually don't have Sakana set here, but I do have an M file. So, I can do a source.env in this directory and that'll pick up uh the Sakana API key for me because it's set in that file. And then what I do is a curl um against the completions endpoint. And uh I'm piping this out to jq, so we'll get pretty JSON. Um and there. I said, "Hey, how many R's are in the word strawberry?" There are three R's in the

9:39word strawberry. And this is the example in the Sakana Docs. They do exactly the same thing. I actually just copied this. So, you can basically the cool thing actually the cool thing here that that I kind of buried the lead on is that we're using the OpenAI API to do this. So, Fugu is

9:56built around the OpenAI responses API. You can just use it's a OpenAI compatible endpoint. We can just call Sakana directly through this endpoint in Python as they're doing here or using curl as we just did. Um this ex- this example. So, only thing you have to do is make sure to set your Sakana API key. Good way of doing that is in a .env file. For Codex it's in your root root Codex directory .env. Boom. Um or by provisioning an API key through a provider like Open Co- Open Router rather and then using Sakana in Open Co. So, quick

10:35video. That's what Sakana Fugu is. That's how you can use Sakana Fugu with Codex. It's how you can use Sakana Fugu with Open Co and it's how you can use Sakana Fugu with an OpenAI compatible API either via the command line or in Python or something. Hopefully, this is all you need to know about Sakana Fugu. It's not demystified. You know, you don't have to go around wondering what is this all powerful model. Um but I'm

10:58Matt. Until next time. Peace.