While Dario Amodei and his team of geniuses have been busy untangling the regulatory mess gifted to them by the US authorities, things are brewing in Japan’s AI industry. Tokyo’s Sakana AI – a startup – has created something that’s essentially as capable as the best that the West has to offer. Meet Fugu – Japan’s take on the likes of Claude Mythos and the conventional mighty AI models in general.
Fugu is a unified multi-agent orchestration system designed to achieve frontier-level AI performance. Sakana, however, has built an agile AI orchestra instead of building another monolithic mega-model. Unlike the models of the West, this one is all about a multi-model orchestrator.
Sakana Fugu: Is it just a Japanese Claude Fable 5?
Theoretically, Sakana’s Fugu could be better for productivity. Let’s dive deeper into what this model is all about.
Anthropic’s Claude Fable 5 is an absolute powerhouse, described by early developers as “relentlessly proactive.” Fable 5 features an always-on “adaptive thinking” loop, allowing it to autonomously spin up testing harnesses, execute complex code over days at a time, and use advanced vision to critique its own work.
Fable 5, though, is massive, computationally expensive ($10 per million input / $50 per million output tokens), and, at the moment, politically vulnerable. The US Department of Commerce imposed export restrictions on the AI model over concerns of high-profile jailbreak situations.
This is where Sakana Fugu takes a radically different approach.
Fugu does not try to be a singular, omniscient mind like Mythos. Behind its single API sits an intelligent router that acts like a master chef, or an orchestral conductor. When a query comes in, Fugu dynamically assesses the problem, builds a custom team of specialised, smaller model agents on the fly, delegates sub-tasks (planning, coding, verification), and synthesises the final output.
“Sakana Fugu is itself an LLM, trained to call various LLMs in an agent pool, including instances of itself recursively. Fugu dynamically orchestrates the world’s best models to tackle complex, multi-step tasks,” says its creator.
The results speak for themselves – Fugu clocked a staggering 73.7 on the SWE-Bench Pro coding benchmark, which is only a little behind Claude Mythos 5’s 77.8% and Claude Fable 5’s 80.3%.
How they differ
Here is how they stack up across key categories:
Core architecture
Claude Fable 5 is built as a monolithic, massive frontier LLM featuring a native, continuous “adaptive thinking” loop. On the other hand, Sakana Fugu is built as a multi-agent orchestration framework utilising a dynamic, swappable pool of smaller, specialised models.
Data retention and privacy
Claude Fable 5 enforces a mandatory 30-day data retention window with no zero-data options due to built-in safety classifiers. Sakana Fugu, meanwhile, offers highly customisable privacy controls, allowing users to route data exclusively to zero-retention, local, or regional models.
Geopolitical risk
With Claude Fable 5, there’s high geopolitical risk, as it is heavily tethered to United States export controls, compliance mandates, and sudden policy shifts.
Sakana Fugu, on the other hand, has a lower risk. It is explicitly engineered for “AI Sovereignty” to allow organisations to bypass vendor lock-in and trade friction.
Key strengths:
Anthropic’s Claude Fable 5 delivers unrivalled single-pass correctness and deep reasoning on ambiguous, long-horizon tasks.
Sakana Fugu, however, provides immense operational agility, total resilience against individual API outages, and highly optimised token economics.
The geopolitical situation gives Japan an AI advantage
Because Claude Fable 5 routes all data through Anthropic’s heavily classified safety layers, it requires a mandatory 30-day data retention window. This has already caused immense corporate friction, causing companies like Microsoft to pull it from internal employee pickers.
Fugu, on the other hand, was built for a fractured geopolitical landscape. Because its underlying pool of model agents is entirely swappable, an enterprise can instruct Fugu to only use locally hosted models or specific European/Asian APIs.
Hence, if a US cloud provider suffers an outage or a sudden regulatory ban, Fugu simply routes around the damage, swapping out the blocked agent for a compliant alternative without breaking the customer’s application.
Who is Sakana Fugu made for?
Similar to Claude Fable 5, Fugu isn’t looking to replace Claude Opus 4.8, Gemini and ChatGPT on an AI hobbyist’s iPhone. It is explicitly engineered as an institutional shield.
– For companies operating in Europe, Asia, or strict regulatory zones, Fugu is an insurance policy. It delivers frontier-level reasoning without binding the company’s fate to American regulatory whiplash or sudden export bans.
– Organisations in finance, healthcare, and defence that legally cannot comply with Western 30-day data logging mandates can use Fugu to orchestrate local and private open-source models under a secure umbrella.
– For teams building heavy agentic workflows, Fugu utilises smaller, specialised models for the grunt work, routing to heavy-duty models only when necessary, thus lowering the cost per task.
Much like its pufferfish namesake, Fugu is small, remarkably well-adapted, and carries a lethal sting for the legacy monolithic giants like Anthropic, Google and OpenAI. For businesses looking to build resilient and sovereign AI systems that no government can switch off overnight, Fugu could impart a new hope.