From Single Giants to Swarms: What Fugu Ultra Actually Is
Sakana AI’s Fugu Ultra is a multi-agent orchestration system that appears as a single AI model but internally breaks complex requests into subtasks, routes them to a pool of expert models, and then recombines their answers to reach frontier-level performance on coding, reasoning, and scientific work.
That design choice is the real story: Fugu Ultra is not another all-purpose giant, it is a conductor. Sakana AI describes it as a language model that can autonomously delegate subtasks to a pool of expert LLMs, including instances of itself, choosing and combining their outputs for the highest quality results. In benchmarks, this Sakana Fugu system reportedly matches or surpasses leading competitors on coding, reasoning, and scientific tasks. According to Sakana AI, “Fugu dynamically orchestrates the world’s best models to tackle complex, multi-step tasks,” positioning itself as a frontier model alternative without owning the biggest model in the room.
The implication is blunt: if multi-agent orchestration works, you do not need to outspend frontier labs on a single enormous model; you need to outsmart them on AI model routing.
How Multi-Agent Orchestration Levels the Playing Field
Fugu Ultra’s bet is that coordination beats raw scale for many real workloads. Instead of a monolithic model trying to be good at everything, Fugu breaks up user prompts into subtasks and determines which subtask to send to which model. That is AI model routing with real teeth, not a thin router that sprays prompts at several models and compares their outputs.
Sakana says Fugu performs as well as Anthropic’s Fable 5 and Mythos Preview on engineering, scientific, and reasoning benchmarks by strategically routing across a swappable pool of expert agents. Because those agents are “entirely swappable,” smaller labs or startups can, in theory, plug in new open or commercial models and still hit near-frontier performance without training Mythos-scale systems themselves.
This is a direct challenge to the orthodoxy that only frontier labs can deliver top performance. If the orchestration layer is smart enough, specialized models coordinated through a Sakana Fugu system start to look like a single, frontier model alternative—especially on multi-step workflows that benefit from different experts handling different steps.
Sovereignty, APIs, and the Risks of Betting on One Giant
The timing of Fugu’s launch is not accidental. An export control directive forced Anthropic to pull Fable 5 and Mythos 5 only three days after launch, a stark reminder that access to top models can vanish overnight. In response, Sakana argues that relying on a single company’s model for critical infrastructure is a massive risk and pitches Fugu as the antidote to single-provider dependence.
From the outside, Fugu looks like one model, exposed through a single OpenAI-compatible API that is available to the general public worldwide. Under the hood, it can route work away from any provider that suddenly restricts access, shifting tasks to other agents in its pool. Sakana AI positions this as a step toward AI sovereignty, promising ongoing improvements as new models are added over time.
Yet the sovereignty claim has limits. Fugu still depends on upstream models; if several providers restrict access simultaneously, Fugu’s capacity shrinks too. Orchestration reduces single-vendor risk; it does not erase systemic dependence.
Theory vs. Reality: Pricing, Speed, and Everyday Use
On paper, Fugu Ultra is a frontier model alternative. In practice, the early user story is far less tidy. Some call it “just a highly advanced router/wrapper, not a fundamental leap like Mythos/Fable,” pushing back on the marketing around collective intelligence. Others report that while Fugu can be “quite strong” for certain agentic coding tasks, it makes mistakes they rarely see from top frontier models.
Pricing and performance are sore points. Fugu and Fugu Ultra subscriptions are offered at USD 20 (approx. RM92), USD 100 (approx. RM460), and USD 200 (approx. RM920) monthly, with pay-as-you-go pricing where Fugu Ultra runs at USD 5 (approx. RM23) per million input tokens and USD 30 (approx. RM138) per million output tokens, rising when context exceeds 272k. Several early users complain about high price-to-burn-rate ratios and APIs they describe as “extremely slow.”
One developer outside the US sums up the gap between promise and practice: they view alternatives to OpenAI and Anthropic as vital, but say Fugu is “nowhere remotely near usable as a day-to-day workhorse.” That tension—impressive demos versus mixed daily experience—will decide whether multi-agent orchestration moves beyond hype.
What Fugu Signals About the Next Wave of AI Strategy
Fugu is less important as a product than as a direction. Sakana was founded with a mission to build collaborative AI ecosystems rather than monolithic models, and CEO David Ha argues that large-scale, single models have had their moment; solving more complex problems will require collective intelligence. Fugu is their first serious attempt to turn that philosophy into infrastructure.
Benchmarks that claim Fugu Ultra matches or beats Gemini 3.1, Opus 4.8, and GPT 5.5 on coding, reasoning, and scientific tasks show how far orchestration can go when tuned carefully. At the same time, mixed user reviews and concerns about price, speed, and burn rate highlight how fragile multi-agent systems can feel once they leave the lab.
The takeaway: multi-agent orchestration will not replace frontier models overnight, but it gives smaller players a path to compete without building one giant system. The winners in the next AI wave may not be whoever trains the biggest model, but whoever designs the smartest conductor—and Fugu is an early, imperfect, but important proof of that shift.
