Meet Fugu: The AI Model That Doesn't Answer Your Question, It Hires a Team
Instead of betting on one giant model, Sakana Fugu dynamically assembles a team of specialist models behind a single API - and says its Ultra tier now competes with frontier systems on coding, science, and reasoning.
What if the next frontier model is not a bigger brain, but a better manager?
That is the bet behind Sakana Fugu, a new product from Japanese AI company Sakana AI. Fugu does not primarily try to answer every prompt itself. It decides which underlying models should work on a task, how to split the work, when to verify the result, and when to call another agent - including itself - recursively.
To the developer, it looks like one model API.
Behind the API, it is an orchestra.
Sakana calls this a “multi-agent system as a model.” Its pitch is straightforward: instead of forcing a single general-purpose model to be brilliant at everything, dynamically assemble a team of models with different strengths for each complex task.
On June 22, Sakana announced two tiers: Fugu, optimized for interactive use and lower latency, and Fugu Ultra, which takes a deeper, quality-first approach through a larger expert-agent pool.
The company says Fugu Ultra stands alongside systems such as Fable 5 and Mythos Preview on high-end coding, reasoning, and scientific benchmarks. That is a big claim. It is also a claim from Sakana’s own release, not an independent benchmark review.
One API, Many Brains
The important distinction is that Fugu is not simply a routing rule written by an engineer.
Sakana says Fugu itself is a language model trained to call models in an agent pool. Given a task, it decides which model to use, whether to delegate planning or execution, whether to ask another agent to check the answer, and how to combine the results.
This is a different product philosophy from the race to train one ever-larger monolithic model.
Sakana’s underlying research points to two papers from ICLR 2026. TRINITY explores evolutionary coordination across Thinker, Worker, and Verifier roles. Conductor uses reinforcement learning to discover communication and collaboration strategies between agents.
The point is not merely to have several agents in a pipeline. It is to learn the pipeline itself.
In Sakana’s framing, Fugu is an AI project manager that can hire a mathematician, a coder, a researcher, and a reviewer for a task - then decide when one of them needs a second opinion.
The Benchmark Claim
Sakana’s release includes a broad comparison across coding, reasoning, and science benchmarks.
For Fugu Ultra, the company reports:
73.7 on SWE-Bench Pro
93.2 on LiveCodeBench
95.5 on GPQA-D
50.0 on Humanity’s Last Exam (text)
82.1 on TerminalBench 2.1
The figures are attention-grabbing, particularly because the chart compares Fugu Ultra with Fable 5, Mythos Preview, GPT-5.5, Gemini 3.1 Pro, and Opus 4.8.
But benchmark tables are never the whole story. They depend on harnesses, budgets, tools, model versions, prompts, and evaluation conditions. The useful takeaway is not that one screenshot settles the frontier race. It is that orchestration has reached the point where it can plausibly compete with monolithic systems on the benchmarks people actually watch.
Six Demos, One Argument
Sakana’s product launch tries to make the benchmark case more concrete with six demonstrations.
In automated ML research, it says Fugu Ultra ran 123 experiments and found the best reported BPB score in its evaluation. In Japanese historical-document recognition, it reports a normalized edit distance of 0.80, well above the cited competitors. In a Rubik’s Cube coding task, it says the generated solver completed all 300 test cases.
The company also shows a CAD iris mechanism, blind chess, and a 50-week trading backtest. As with the benchmark table, these are company-reported demonstrations rather than outside reproductions. Still, they illustrate the product thesis: Fugu is meant for messy tasks where planning, verification, iteration, and specialist handoffs matter more than a single clean prompt.
The Geopolitical Pitch Is Not Subtle
Sakana repeatedly frames Fugu as offering frontier capability without single-vendor dependency or export-control risk.
That is not a technical footnote. It is the commercial argument.
If one underlying provider becomes unavailable because of policy, regulation, price, or an outage, an orchestration layer can route work elsewhere. Performance may change, but the application does not need to go completely dark.
For companies building AI into internal workflows, that flexibility can be more valuable than squeezing out a few extra benchmark points.
Fugu also allows customers to opt out of certain providers or models in the standard tier. Ultra uses a fixed agent pool, according to Sakana.
Pricing and Availability
Sakana offers subscriptions at $20/month, $100/month, and $200/month, covering both Fugu and Fugu Ultra with different usage limits.
For API use, the company lists Fugu Ultra at $5 per million input tokens and $30 per million output tokens, with higher prices beyond 272K context. It says the API is OpenAI-compatible, so developers can switch by changing the endpoint rather than adopting a proprietary SDK.
Fugu is currently unavailable in the EU and EEA while Sakana works toward GDPR and region-specific compliance.
Why This Matters
There are two ways to see Fugu.
The cynical view is that it is a premium model router with a very good marketing story.
The more interesting view is that model orchestration is becoming a new scaling axis. As the raw capabilities of frontier models converge, the advantage may increasingly come from how well a system selects models, manages context, checks its own work, recovers from failure, and adapts to a user’s privacy or vendor constraints.
Fugu’s real test will be whether developers find it meaningfully better on long, ambiguous, high-stakes work than calling a single flagship model directly.
But Sakana has made the strategic direction unmistakable: the next frontier AI may not be one model at all.
It may be a model that knows how to run a team.