From One Big Brain To A Team Of Specialists
Sakana AI’s Fugu and Fugu Ultra are multi-agent orchestration systems that act as a single language model while internally routing subtasks across a pool of specialized expert models to match frontier-level AI performance on complex, multi-step workflows. This is not another giant monolithic model trying to beat everyone at scale; it is a coordinated crew. Fugu itself is trained for delegation, inter-agent communication, and result aggregation, turning a model routing system into an active conductor of work rather than a passive switchboard. The release of Fugu on June 22, 2026, alongside the frontier-level Fugu Ultra, makes this architecture available worldwide through one OpenAI-compatible API, targeting demanding use cases in engineering, science, research, cybersecurity, and data analysis. The message is clear: if you cannot own the biggest model, own the smartest team.
How Fugu Ultra Turns Routing Into Collective Intelligence
Fugu Ultra distinguishes itself by using multi-agent orchestration: it is a language model that autonomously breaks down user prompts, delegates subtasks to expert LLMs, including instances of itself, then verifies and synthesizes their outputs. Unlike simple multi-model routers that spray the same prompt across several models and blend results, Fugu behaves more like a project lead. It decides which subtask goes to which specialized expert model and when its own capabilities are enough, acting as model selector, delegator, verifier, and synthesizer within one interface. According to Sakana AI, “Fugu dynamically orchestrates the world’s best models for complex, multi-step tasks,” delivered as a single OpenAI-compatible endpoint that hides the underlying complexity from developers. Architecturally, this is a bet that learned coordination can turn a collection of good models into a practical frontier AI alternative.
Benchmarks, Pricing, And The Trade-Offs Of Going Agentic
On paper, Fugu Ultra’s numbers put it in frontier territory: it scores 73.7 on SWE-Bench Pro, ahead of Claude Opus 4.8’s 69.2 and GPT-5.5’s 58.6, and reaches 50.0 on Humanity’s Last Exam, essentially matching Opus 4.8’s 49.8. Qualitative reports back this up, with users seeing stronger sustained persona consistency and more thorough issue discovery during long code review and security assessment runs. The practical impact is real: an industry researcher completed patent landscape analysis across roughly 20 papers and several patents in a few hours instead of three to four days. But early adopters are not blind to the costs. Fugu Ultra has fixed pricing at USD 5 (approx. RM23) per million input tokens and USD 30 (approx. RM138) per million output tokens, with monthly tiers at USD 20 (approx. RM92), USD 100 (approx. RM460), and USD 200 (approx. RM920), and some users complain about “extremely slow” API calls and burn rates that escalate too quickly.
Sovereignty Claims And The Limits Of Swappable Agents
Sakana AI frames Fugu as a step toward AI sovereignty, arguing that relying on a single company’s frontier model for national infrastructure is a massive risk, especially after export controls forced Anthropic to pull Fable 5 and Mythos 5 days after launch. By coordinating a pool of entirely swappable agents, and letting enterprise teams choose which models participate to satisfy data residency or compliance rules, Fugu promises resilience: if one provider disappears, the system can route work to others. Yet this sovereignty story has cracks. Fugu still relies on other vendors’ models, and some early users question whether this dependence, combined with high prices and slow inference, delivers more practical freedom than a well-negotiated contract with a single frontier provider. Critics argue that while Fugu is more than a router, its multi-agent orchestration does not fully escape the geopolitics and supply constraints of the underlying model ecosystem.
A Different Competitive Play For Frontier AI Alternatives
The real significance of Fugu Ultra is strategic: it shows how a startup valued at USD 2.65 billion (approx. RM12.2 billion) can compete with frontier labs without winning the scale race. Instead of training a single all-purpose giant, Sakana coordinates a pool of existing specialized expert models and wraps them in a learned orchestration layer that feels like one model via an OpenAI-compatible API. This approach addresses a genuine deployment problem: complex tasks need different models at different stages, and Fugu reduces integration friction by automating that choice and aggregation. It will not be the blueprint for sovereignty many hope for, and its price–performance–latency mix will deter some teams. But as multi-agent orchestration matures, Fugu makes a strong case that the next wave of frontier AI alternatives could come less from bigger brains and more from smarter coordination.
