Sakana AI launched a product on Monday that takes a fundamentally different approach to frontier AI capability. Rather than training a single massive model to do everything, Fugu coordinates a team of specialized AI systems — routing tasks by complexity, dynamically assigning Thinker, Worker, and Verifier roles, and synthesizing the outputs into a single response.
The Tokyo research lab claims the result matches Anthropic‘s Fable 5 and OpenAI‘s GPT-5.5 on major benchmarks. And it launched ten days after US export controls restricted access to both those models for much of the world.
Read: Anthropic Suspends Fable 5 Access After US Government Export Control Order
Sponsored · Cambria Gold Mines
Most AI products funnel everything through one powerful model. That approach has two problems, with both getting harder to ignore.
First, different models genuinely outperform each other on different types of tasks — a model that excels at code review may be weaker at scientific reasoning. Second, access to the best models has become geopolitically uncertain. On June 12, the US government imposed export controls on Anthropic’s Fable 5 and Mythos Preview, restricting which countries could use them. For any organization that had built critical infrastructure on those models, access could disappear based on a foreign policy decision outside its control.
Fugu is Sakana’s answer to both vulnerabilities at once.
What Fugu actually does
Fugu is not a traditional large language model trained on massive compute. It is a trained orchestration model — a system that has learned to coordinate a team of other AI models rather than do everything itself.
From a user’s perspective, Fugu looks identical to any other AI model. A single OpenAI-compatible API call goes in. Behind the scenes, Fugu’s conductor model reads the task and decides how to handle it — answering directly if the task is simple, or assembling a coordinated team of specialist models if the problem is complex.
Sponsored · Selkirk Copper Mines
Thinker, Worker, and Verifier roles are assigned dynamically across multiple frontier models, which communicate and refine their outputs. Fugu then synthesizes the result into a single coherent response. The entire coordination layer is invisible to the developer’s code.
The key distinction from existing orchestration frameworks like LangGraph or CrewAI is that those tools require developers to write the routing logic themselves. Fugu learns the routing. The technical foundation comes from two ICLR 2026 papers Sakana published earlier this year: TRINITY, an evolved LLM coordinator, and The Conductor, which trains orchestration strategies through reinforcement learning.
Unlike most agent frameworks, the coordination approach here has a published, peer-reviewed academic foundation.
Two tiers
Sakana ships Fugu in two versions. The standard Fugu targets everyday tasks — coding, chat, document review — optimizing for low latency. Fugu Ultra handles demanding multi-step problems: security assessments, Kaggle competitions, and extended research workflows.
Ultra is priced at $5 per million input tokens and $30 per million output tokens, with rates doubling for contexts above 272,000 tokens.
The benchmark claims — and their limits
Sakana claims Fugu Ultra matches Fable 5 and GPT-5.5 across engineering, scientific, and reasoning benchmarks — scoring 93.2 on LiveCodeBench against Fable 5’s 89.8.
However, there are three caveats: 1) these benchmarks are self-reported by Sakana and have not been independently verified; 2) neither Fable 5 nor Mythos Preview is in Fugu’s agent pool — export controls prevent it (as one independent review put it, “matching them is therefore a claim about substitutes, not a way to get their output”); and 3) whether reselling access to third-party models through a single API complies with each provider’s terms of service is an open question — one every enterprise customer inherits.
The Japan angle
Japan’s Digital Minister had warned earlier this month that without faster domestic AI development, the country risked becoming an “AI colony.” Sakana’s launch — ten days after US export controls restricted access to Fable 5 and Mythos Preview — was a direct answer to that concern.
“Recent disruptions in the AI landscape have demonstrated the severe risk of single-vendor dependency,” the company wrote, describing reliance on a single provider’s APIs for critical infrastructure as a vulnerability that had moved from hypothetical to real.
Fugu also has a structural advantage that grows over time — every new capable model added to its pool makes the whole system stronger, without requiring Sakana to retrain anything.
