Sakana Fugu, Arrives. The Answer to the "Single Dependency" Problem from the Commander Model Orchestrating Multiple AIs

Wait a second, isn't June 12th still fresh in your memory?

Three days after Anthropic released its top-tier model, "Claude Fable 5," access was abruptly cut off. Due to export regulations, it became unavailable from certain countries. I'm sure I wasn't the only one who thought, "Suddenly?" and it caused quite a stir in overseas communities as well.

I think that event brought one thing into sharp focus.Relying entirely on a specific company's API is, in itself, a risk.

And, as if timed to match that moment, this service from Tokyo appeared.

On June 22, 2026, Sakana AI released "Sakana Fugu" to the public. In short, it's a system that "bundles multiple AIs internally and allows them to be used as a single model from the outside," but it has a more interesting structure than I expected.

🐡 What does Fugu actually do?

First, let's organize the basic mechanism.

Typical AI services work like this: If you want to use Claude, you hit the Anthropic API. If you want to use GPT, you hit OpenAI. We, the users, decide which model to call.

Fugu is designed the opposite way.

Users simply send requests to a single endpoint called Fugu. From there, Fugu determines the difficulty and type of the task and autonomously selects and operates the models.

Moreover, the way Fugu decides who to delegate to is unique; it's not based on "hand-written rules." Fugu itself learns "which combination of models produces the best answer." Internally, it dynamically assigns "thinker," "executor," and "verifier" roles to multiple models depending on the task to advance the processing.

To use a human work analogy, it's like a talented project manager who receives a project and organizes the team themselves, saying, "I'll leave this to Person A, and have Person B check this."

The foundation of this research is two papers that Sakana AI presented at ICLR 2026.

• TRINITY — A lightweight coordinator that dynamically assigns "Thinker / Worker / Verifier" roles to multiple models

• Conductor — A mechanism that uses reinforcement learning to have agents learn the coordination strategy itself

It's an approach where the AI discovers "how to conduct well" on its own, rather than using hand-designed workflows. It feels very Sakana-like, doesn't it?

🔬 So, how strong is it really?

Every time a service like this comes out, the thing I'm curious about is the benchmark. According to Sakana's announcement, the higher-end version, Fugu Ultra, consistently outperformed Gemini 3.1 Pro, GPT 5.5, and Claude Opus 4.8 on benchmarks like SWE-Bench Pro, GPQA-D, and Humanity's Last Exam.

I want to add a proper note here. The benchmarks are based on Sakana's self-reporting. Fugu Ultra's agent pool does not include Fable 5 and Mythos Preview, which are currently inaccessible. We are in the phase where independent verification by third parties is yet to come.

However, there is one interesting thing.

Fugu is not a single base model, but a coordinator model that has learned 'how to bundle' multiple models. Therefore, the underlying idea is that, in principle, it has the potential to achieve higher performance than each of the individual models in the pool. Isn't it interesting that this is a different approach from 'trying to build one amazing model'?

There are several examples published as actual test cases.

In a task to solve a Rubik's Cube, Fugu Ultra completed it in 19.72 moves (the difference from GPT-5.5 was almost negligible), while two models crashed and failed to solve it. In a reading test for ancient Japanese documents, Fugu Ultra recorded 0.80 in NED (an index that is the reciprocal of character error rate), while the next best model was 0.24. The difference in old Japanese documents is quite significant.

💰 What is the cost if I use it?

Two systems are available: subscription and pay-as-you-go.

For subscriptions, Standard is $20/month, Pro is $100/month, and Max is $200/month, and both Fugu and Fugu Ultra can be used in all plans. My impression is that the price range is about the same as major competing services.

The API pay-as-you-go Fugu Ultra (fugu-ultra-20260615) has a two-tier system: $5 for input and $30 for output per 1 million tokens, with cached input at $0.50. When the context exceeds 272K tokens, it increases to $10 for input and $45 for output, so you will need to estimate costs for tasks like reading an entire codebase.

For now, it seems there is a campaign where if you register by the end of July 2026, the second month is free. Please check the official website for details.

🧭 The context of 'Tokyo-based'

For those who are wondering, who is Sakana AI anyway? A brief explanation.

Founded in Tokyo in 2023. It was started by Llion Jones, a co-author of Google's 'Attention Is All You Need' paper, and David Ha, former head of research at Stability AI. The philosophy of 'having multiple models cooperate rather than building one large model' has been consistent from the beginning, and the company name 'sakana' (fish) comes from that metaphor.

In a funding round in late 2025, the valuation reached $2.6 billion, and in June of this year, they also released the autonomous research agent 'Marlin' for B2B. Fugu was released right after that, symbolizing the transition from the research phase to the commercial product phase.

David Ha himself said this at the time of Fugu's release: that relying on a single company's API for critical infrastructure, finance, and governance is 'no longer a hypothesis but a real risk.' Considering the context of the release the week after the Fable access outage, I don't think the timing is a coincidence.

🤔 Things I'm curious about

To be honest, there are parts of Fugu that I 'still don't quite understand'.

Routing is a black box, and it is designed so that 'which model actually ran for the current request' is not disclosed. From an enterprise governance perspective, opinions on how to evaluate this will likely vary.

Another point is that dynamically calling multiple frontier models means that costs can be difficult to predict. For simple requests, Fugu itself answers and it's cheap, but if you call frontier models in multiple stages for difficult tasks, you won't know the actual cost until you use it.

I understand the design philosophy that 'it's cheaper in total because it uses a smarter approach than calling the most expensive model directly,' but you can't judge whether that applies to your own use case without trying it.

💡 A perspective I find interesting

What I find most interesting about Fugu is that it presents a different axis from the arms race of 'building one smart model'.

The idea that 'instead of raising a smart model, if we learn a mechanism to skillfully bundle existing models, we might be able to surpass any single model.' And the commercialization of that as 'providing it as a model via API.'

To use a gaming analogy, it's like creating a commander unit that has learned 'how to position and deploy existing strong units to fight' rather than raising one single strong unit. This concept itself is quite characteristic of Sakana.

Furthermore, Japanese companies, especially in the financial sector, are actually moving quickly. News has emerged that Sumitomo Mitsui Banking Corporation is automating proposal creation with multi-agents, and MUFG is utilizing it to support the creation of loan approval documents, showing that Sakana's multi-agent technology is penetrating heavy industries in Japan.

Summary

To summarize Sakana Fugu in one phrase, it is a 'commander model that bundles multiple AIs.' It is a new approach that has learned a mechanism to coordinate multiple existing models rather than building one large model.

Performance evaluation is just getting started, so there is no need to take it at face value, but I find the timing of the release in terms of the 'risk of dependency on a single vendor' and the fact that such a design emerged from Tokyo to be interesting.

If both Fugu and Fugu Ultra can be accessed from a $20/month plan, there is surely enough reason to give it a try. What kind of tasks would you like to use it for? Let me know in the comments.

Production Notes

This time, I structured the article with the context that the recent news of 'Fable 5 access suspension' and the release of Fugu are not coincidental but connected. Rather than writing it as a technical explanation article, I wanted to read it as 'Fugu demonstrates what that event signifies.'

I included the issues of routing black boxes and cost readability because I thought it was more like Miccell to write about them frankly rather than hiding them just to make the article read positively.

Miccell - Once you understand how things work, the world becomes more interesting.