Fugu Ultra in One Sentence: Frontier Power by Multi-Agent Design
The Sakana Fugu system is a multi-agent orchestration layer that presents as a single language model but internally breaks complex requests into subtasks, routes them to a pool of specialized expert models, and then verifies and synthesizes the results to match frontier-level performance on coding, reasoning, and scientific work. This week, Sakana AI released Fugu and its flagship Fugu Ultra as frontier-level orchestration models that aim to rival Anthropic’s Fable 5 and Mythos Preview while reducing the risk of depending on a single AI provider. The key idea is blunt: instead of spending billions to train one huge model, a smaller startup coordinates many strong models through an orchestration API and competes on how intelligently it routes and combines their outputs.
How Multi-Agent Orchestration Tries to Match Frontier AI
Fugu Ultra is not marketed as “yet another router” for good reason. It is itself a language model, specialized for model selection, delegation, verification, and synthesis; it can answer requests directly or decompose them into subtasks and assign each subtask to an expert agent from its swappable pool, including instances of itself. Unlike multi-model routers that broadcast one prompt to several models and then compare or blend the outputs, Fugu performs task-specific AI model routing: different parts of a workflow can hit different models, and only the final result is exposed through what looks like a single OpenAI-compatible orchestration API. Benchmarks cited by Sakana claim that Fugu Ultra “matches or surpasses leading competitors” on coding, reasoning, and scientific tasks, and that it consistently beats Gemini 3.1, Opus 4.8, and GPT 5.5 on selected tests.
Accessibility Without Monoliths—and the Real Cost to Developers
Strategically, Sakana is making a clear pivot: build collaborative AI ecosystems instead of monolithic models, and expose them through an OpenAI-compatible API so developers can adopt Fugu with minimal friction and less vendor lock-in risk. From the outside, Fugu looks like any other single frontier model: one endpoint, same client libraries, no need to understand the internal model roster. In practice, the company offers two tiers—Fugu as a lower-latency daily driver and Fugu Ultra as the heavy orchestration engine for complex engineering and research workflows. That accessibility comes at a price. Subscription plans start at USD 20 (approx. RM94) and go up to USD 200 (approx. RM940) per month. Fugu Ultra’s pay-as-you-go rates of USD 5 (approx. RM23) per million input tokens and USD 30 (approx. RM140) per million output tokens, with higher costs beyond a 272k context, have already triggered complaints about fast burn rates and high effective costs for long-running tasks.
Sovereignty Claims Meet the Messy Reality of Multi-Provider Dependence
Sakana frames Fugu Ultra as a step toward AI sovereignty, arguing that relying on a single company’s model for critical infrastructure is a massive risk in a world where export controls can remove frontier access overnight. The jab is obvious: Anthropic had to pull Fable 5 and Mythos 5 only days after launch due to an export directive, confirming that access to one vendor’s frontier AI can vanish without warning. Fugu’s answer is “collective intelligence” built from an entirely swappable pool of agents; if one provider restricts access, in theory tasks can be rerouted to another. But critics are right to say this is not a real blueprint for sovereignty. If multiple providers clamp down, Fugu’s capabilities shrink too. Its orchestration layer reduces single-provider fragility, yet it cannot escape the upstream dependencies it relies on. The system is more than a router, but it is still a wrapper around other people’s models.
Early User Sentiment: Strong Idea, Uneven Execution
On paper, Fugu Ultra looks impressive: almost 500 beta users reportedly tested it on lengthy multi-step workflows, and some cybersecurity teams praised its tendency to stay within safe parameters and surface more issues than competing models in code review and security assessment. Users mention “unusually strong persona stability” across long sessions, something many single-model systems still struggle with. Yet early public feedback is mixed. Some developers describe the tool as quite strong for agentic coding but weaker on implementation details where frontier models rarely slip. Others slam the price-to-burn-rate ratio, saying the API feels extremely slow and that real-world quality falls short of direct access to Fable-level systems. One quotable summary from a skeptical developer: “It’s nowhere remotely near usable as a day-to-day workhorse”. The concept of multi-agent orchestration is convincing; Fugu’s current execution is not yet equally convincing.
Conclusion: Orchestration Will Matter, But Fugu Must Prove It at Scale
Fugu Ultra shows how smaller AI companies can compete with frontier AI rivals by coordinating many expert models instead of training one massive system. Multi-agent orchestration and task-specific AI model routing are likely to become standard for complex workflows; they make far more sense than forcing a single model to be good at everything. But an orchestration API is only valuable if it feels fast, affordable, and reliably better than calling frontier models directly. Today, Fugu is an ambitious, opinionated bet on collective intelligence and multi-provider resilience, wrapped in an appealing OpenAI-compatible interface. Tomorrow, its fate depends on whether Sakana can tame burn rates, improve latency, and keep adding strong expert agents. Orchestration may win the long game—but Sakana still has to prove that Fugu, not a frontier lab, will be the one orchestrating it.
