Sakana Fugu is not a new AI model. It’s not a framework. It’s not a chatbot. A Japanese lab just launched something different: a coordinator. You send it a task, it decides which AI models to call, in what order, and how to stitch the results together. You just get an answer.
No picking the right model. No juggling APIs. No wondering if Claude would have done better than GPT on that last task.
That’s the pitch from Sakana AI, a Tokyo-based research lab that dropped Fugu into general availability today. And the timing is not an accident.
What Sakana Fugu Actually Does
Most AI tools give you one model and hope it’s good enough. Sakana Fugu works differently.
Under the hood, Fugu is itself a language model, but one trained specifically to coordinate other AI models. When you send it a request, it breaks the task down, assigns pieces to different models in its pool, verifies the results, and combines everything into a single response. From your side, there’s one API endpoint. The complexity stays invisible.
Sakana Fugu is a multi-agent orchestration system that behaves like a single foundation model.
Think of it like a project manager who has a team of specialists. You don’t talk to the specialists directly. You talk to the manager, and they handle the delegation. Except in this case, the manager is also an AI, and the specialists are other frontier models working behind the scenes.
There are two versions at launch. Fugu handles everyday tasks with lower latency, good for coding tools, review workflows, and interactive use. Fugu Ultra is tuned for maximum accuracy on demanding, multi-step problems, and is where the headline benchmark claims live.
One thing Sakana has not disclosed is which specific models sit inside the pool. They’ve confirmed that Claude Fable 5 and Mythos Preview are not included, because those models are currently export-controlled and publicly inaccessible. Beyond that, the pool composition is not public. Based on the benchmark comparisons published at launch, Opus 4.8, GPT-5.5, and Gemini 3.1 Pro are the models Fugu is measured against, which suggests they are likely candidates. But that’s an inference, not a confirmed list.
The Export Control Angle Nobody Is Talking About Enough
Here’s where this gets genuinely interesting for regular people, not just developers.
On June 12, 2026, some of Anthropic’s most capable models became subject to national-security-based export controls. Overnight, organisations in a broad set of countries lost access to tools they had built critical workflows around. No warning. No migration path. Just gone.
Sakana launched Fugu ten days later with a deliberately pointed message: frontier AI capability, without the risk of export controls.
Their CEO David Ha put it plainly in the launch post. Relying on a single company’s APIs for critical infrastructure, finance, or governance is a material vulnerability. That risk is no longer hypothetical. It already happened.
Fugu’s agent pool is designed to be swappable. If one provider restricts access, the system routes around it. As new models enter the pool, including Sakana’s own models over time, users get the benefit automatically without changing a line of code.
That’s a genuinely different value proposition from anything else on the market right now.
How to Actually Use It
If you already call any AI via API, switching to Fugu requires almost no work. Fugu uses an OpenAI-compatible API, which means you point your existing integration to a new endpoint and change a few lines of configuration. That’s it. No new SDK to learn, no major refactor.
For people who don’t work with APIs at all, there’s a chat interface available at chat.sakana.ai where you can use it directly in a browser, similar to how you’d use ChatGPT or Claude.ai.
Pricing sits at roughly $20 to $200 per month depending on your usage tier, with a pay-as-you-go option for heavier workloads. That puts it in the same ballpark as most major AI subscriptions. One detail worth noting: even when multiple agents run simultaneously on your request, billing does not stack. You’re charged at the rate of the top model used, not the sum of every agent that ran. For complex tasks that would normally require chaining multiple API calls yourself, that could work out cheaper in practice.
Currently not available in the EU or EEA while Sakana works through GDPR compliance. UK users can access it today.
What Early Users Are Saying
Sakana shared a handful of testimonials from users who were part of the early access period. These are Sakana-selected so treat them as illustrative rather than independent reviews, but the use cases they describe are telling.
One software engineer reported that Fugu Ultra surfaced over twenty issues in a code review where other tools typically found three. A cybersecurity analyst said it kept a scoped assessment inside bounds while producing evidence and retest steps throughout the process. An enterprise executive flagged persona stability across long sessions as a standout feature, which matters when you’re building agent products that need to stay on task over extended workflows.
These are the right things to measure. AI agents tend to fall apart in the boring middle of a task: losing scope, dropping context, forgetting constraints. A coordinated system with verification built in should, in theory, be more resilient there. Independent testing will tell us whether that holds.
Frequently Asked Questions
What is Sakana Fugu and how does it work?
Sakana Fugu is a multi-agent orchestration system from Tokyo-based Sakana AI, launched on June 22, 2026. It routes your request through a pool of frontier AI models, coordinates them internally, and returns a single combined response. From the user’s side it behaves like one model. The coordination happens behind the scenes.
How is Sakana Fugu different from just using ChatGPT or Claude?
ChatGPT and Claude are single models. You get one brain per request. Sakana Fugu coordinates multiple models, assigning different parts of a task to different specialists and verifying the output before returning it. The tradeoff is potentially better performance on complex tasks, with less visibility into exactly what ran under the hood.
Which AI models does Sakana Fugu use?
Sakana has not publicly named the specific models in the pool. They’ve confirmed that Fable 5 and Mythos Preview are not included because those models are export-controlled and not publicly accessible. The benchmarks at launch compare Fugu against Opus 4.8, GPT-5.5, and Gemini 3.1 Pro, which suggests those are likely in the mix, but that has not been officially confirmed.
Can I use Sakana Fugu in the UK?
Yes. UK users can access Sakana Fugu at launch. EU and EEA users are excluded for now while Sakana works toward GDPR compliance.
Is Fugu Ultra actually better than Claude or GPT?
Fugu Ultra benchmarks ahead of publicly accessible models like Opus 4.8, GPT-5.5, and Gemini 3.1 Pro on several engineering, reasoning, and agentic tasks. It claims to match Fable 5 and Mythos Preview, though those comparisons are based on vendor-reported scores, not independent testing. Worth watching once third-party evaluations come in.
The Bigger Picture
The most interesting thing about Sakana Fugu is not the benchmarks. It’s the idea underneath them.
The AI industry has spent the last few years in a race to build the biggest single model. Sakana is betting the next phase looks different: smarter coordination between good models, rather than one model trying to do everything alone.
Whether Fugu delivers on that in production is something that will take months to find out. But for anyone who got caught out when access to their AI tools disappeared overnight on June 12, the argument for not putting all your eggs in one API is no longer theoretical.
You can try Sakana Fugu at sakana.ai/fugu.
Leave a Reply