From Monolithic Models To Multi-Agent Orchestration
Multi-agent orchestration in AI is an architectural approach where a coordinating language model breaks complex user requests into subtasks, routes each subtask to specialized expert agent systems, and then verifies and combines their outputs to deliver a single high-quality answer, instead of relying on one monolithic frontier model to do everything end-to-end. Sakana AI’s launch of Fugu and its frontier-level Fugu Ultra puts that idea at the center of the AI race. Rather than trying to outscale Anthropic or other giants, Sakana is betting that smart AI model routing across a swappable pool of experts can match their performance on engineering, science, reasoning, and agent benchmarks. That is a bold claim—and a welcome challenge to the assumption that only massive labs can deliver frontier model alternatives.
How Fugu’s Expert Agent System Works
Fugu Ultra is not presented as a single all-knowing model; it is an orchestrator that decides which expert agents to use for each part of a task. Internally, Fugu is itself a language model specialized for model selection, delegation, verification, and synthesis, meaning it can both solve some tasks directly and break others into subtasks for external models. Unlike simple multi-model routers that spray the same prompt to several models and compare results, Fugu “breaks down user prompts into subtasks and determines which subtask to send to which model.” Sakana says this collective intelligence approach, grounded in its learned orchestration research (Trinity and Conductor), lets Fugu Ultra match or surpass leading competitors on coding, reasoning, and scientific tasks. In other words: intelligence is shifted from sheer parameter count to the quality of AI model routing and task decomposition.
API Access, Cost Efficiency Claims, And Early Friction
On paper, Fugu looks friendly to ordinary developers: from the outside, it behaves like one model, exposed through a single OpenAI-compatible API and generally available in most regions. There are two tiers—a lower-latency Fugu for day-to-day chatbot-style work and Fugu Ultra for deeper, multi-step workflows in engineering, research, cybersecurity, and data analysis. Subscription plans sit at USD 20 (approx. RM92), USD 100 (approx. RM460), and USD 200 (approx. RM920) per month, while Fugu Ultra usage is billed at USD 5 (approx. RM23) per million input tokens and USD 30 (approx. RM138) per million output tokens, with higher rates beyond a 272k context window. Sakana pitches this orchestration as AI cost efficiency: route only what needs heavy firepower to expensive experts, keep everything else lightweight. Early users, however, complain about fast burn rates, high prices, and an “extremely slow” API that can feel like paying more for a slower black box.
AI Sovereignty: Clever Architecture Or Marketing Stretch?
Sakana explicitly connects Fugu to AI sovereignty, arguing that relying on one provider is a “massive risk” now that export controls can cut off access to top models overnight. The sudden withdrawal of Anthropic’s Fable 5 and Mythos 5 after an export directive is the cautionary tale behind this stance. Fugu’s answer is a pool of “entirely swappable agents” that can be reconfigured when a provider tightens access, with ongoing improvements as new models join the pool. Conceptually, this is smart resilience: multi-agent orchestration becomes a hedge against single-vendor lock-in. Practically, critics argue Fugu is still dependent on external models and so cannot be the sovereignty hero it claims. If several providers restrict access at once, Fugu’s capabilities degrade along with them. The architecture limits fragility, but it does not replace the need for open, locally controlled frontier model alternatives.
A Real Shift Beyond Scale—But With Sharp Edges
What matters most about Fugu is not whether it perfectly matches Fable 5 or Mythos on every task; it is that a smaller lab is competing through design rather than raw scale. Sakana was founded to build collaborative AI ecosystems instead of monolithic models, and Fugu Ultra is the clearest expression of that idea so far. Moving forward, Sakana plans to add more models to its agent pool, further refining how it dynamically orchestrates “the world’s best models” for complex workflows. Benchmarks and some user stories show strong persona stability and deeper issue-finding in long code review and security sessions. Yet mixed community sentiment on quality, speed, and price is a reminder that clever architecture does not absolve you from delivering reliable everyday performance. Multi-agent orchestration is now a serious frontier model alternative—but to democratize advanced AI, it must become both technically impressive and financially sane.
