What Is Sakana Fugu? The Japanese AI That Orchestrates an Army of Models (2026)

Last updated: June 2026

In June 2026, a Tokyo startup did something that sounds almost contradictory: it released an AI that performs at frontier level without building a frontier model of its own. Sakana Fugu does not try to be the smartest single brain in the room. Instead, it acts as a conductor, calling on a pool of expert models and coordinating them so that the result feels like one seamless system. For founders, engineers, and AI buyers, that design choice is more than a curiosity. It hints at where competitive advantage in AI may be moving next, from owning the biggest model to owning the layer that orchestrates everyone else's. This guide explains what Fugu actually is, how the orchestration works step by step, how it benchmarks against models like Claude Fable 5, what it costs, and where it genuinely falls short, with honest numbers instead of launch-day hype.

Quick answer: Sakana Fugu is a roughly 7-billion-parameter "conductor" model from Sakana AI in Tokyo. It is not a giant standalone model and not a simple rule-based router. It is a small model trained with reinforcement learning to decide when to answer directly and when to assemble and coordinate a team of other expert models, presenting one OpenAI-compatible API to the user.

What Sakana Fugu actually is

Fugu (フグ, the Japanese pufferfish) was announced for general availability on June 22, 2026, after an earlier beta in April. It comes from Sakana AI, a Tokyo-based research company founded by David Ha (CEO) and Llion Jones (CTO). Jones is one of the co-authors of the 2017 paper "Attention Is All You Need," the work that introduced the Transformer architecture underpinning nearly every modern large language model. That pedigree matters because Fugu is, at heart, a research bet about architecture rather than scale.

The easiest way to misunderstand Fugu is to call it a router. A traditional router is a piece of hand-written logic: if the prompt looks like code, send it to model A; if it looks like a math problem, send it to model B. Fugu is not that. Fugu is itself a trained language model, a compact conductor of around 7 billion parameters, whose entire job is orchestration. It was trained with reinforcement learning, where the reward signal was simple and brutal: was the final answer correct and properly formatted? Over many training rounds, the conductor learned its own strategies for getting that reward, rather than having a human script them in advance.

Sakana describes the result as "a multi-agent system that behaves like a single model." You call one model. On the inside, a coordinated system of experts does the work, and you receive a single coherent answer. The tagline the company uses captures the ambition: "One Model to Command Them All."

There are two variants, and it is worth getting the names right because earlier beta material used different ones. The launch lineup is Fugu (the balanced, lower-latency default) and Fugu Ultra (tuned for maximum answer quality on hard, multi-step problems). An April beta page briefly called the smaller variant "Fugu Mini," but the shipping product uses Fugu and Fugu Ultra. Switching between them is just a model-parameter change in the same API call.

If you want to experiment with the kind of expert models that sit inside orchestration pools, Truescho's free AI tools are a low-stakes place to get a feel for how different models behave before you commit budget to any one provider.

Why Fugu arrives now: the export-controls backdrop

Timing is part of the story. On June 12, 2026, US export controls restricted access to Anthropic's Fable 5 and Mythos models, which are not publicly accessible as a result. Ten days later, Sakana shipped a product whose entire pitch is that you should not depend on any single model, vendor, or jurisdiction. The company's framing is blunt: access to the top AI systems you rely on today "can vanish overnight."

For a business, the relevant lesson is not political and certainly not about working around any law. It is about concentration risk. If a critical workflow, customer support, code generation, document analysis, is wired to exactly one provider, then a price change, an outage, a policy shift, or an export restriction becomes a single point of failure for your operation. Fugu's answer is to make the model layer fungible: orchestrate across many publicly available providers so that if one becomes unavailable, the system routes around it. This is the same logic companies already apply to cloud regions, payment processors, and suppliers. It is vendor diversification, applied to AI.

It is important to be precise here. Fugu does not give you access to Fable 5 or Mythos. Those models are export-controlled and are not in Fugu's pool. What Fugu does is match their benchmark scores by skillfully coordinating other models that you can access. That distinction, matching the frontier without containing it, is the entire commercial proposition, and we will return to whether it holds up.

How the orchestration works, step by step

Underneath the single API call, Fugu runs an invisible four-stage loop: selection, delegation, verification, synthesis. Walking through a concrete example makes it clear.

Imagine you send Fugu Ultra a hard request: "Reproduce the core experiment in this machine learning paper and tell me whether the reported result holds." Here is roughly what happens inside.

1. Selection. The conductor reads the request and decides how to handle it. For something this complex, answering directly would be a poor strategy, so it chooses to assemble a team. Crucially, the conductor learned during training which kinds of problems benefit from teamwork versus a direct answer; this was not hand-coded.
2. Delegation. The conductor splits the work and hands sub-tasks to specialist models in the pool. One model might parse the paper's methodology, another might generate the experimental code, and a third might review that code for correctness. The conductor can also run several models in parallel on the same sub-task and keep the best output.
3. Verification. Rather than trusting any single sub-answer, the conductor checks the work, cross-examining outputs, re-running steps, and discarding results that fail. This is where Fugu's reward training pays off; it was rewarded for correct final answers, so it learned to be skeptical of intermediate ones.
4. Synthesis. Finally, the conductor stitches the verified pieces into one coherent response and returns it to you as if a single model had produced it.

A few learned behaviors are worth highlighting because they emerged from training rather than being programmed. Fugu can route an entire problem to one model when that is enough. It can split a task across several. It can run models in parallel and pick the strongest result. And it can call itself recursively, effectively spending more compute at inference time to push quality higher on the hardest problems, with no retraining required.

Two practical properties matter for businesses. First, the agent pool is fully swappable; if a model becomes unavailable, Fugu dynamically routes around the disruption. Second, organizations can exclude specific agents or providers from the pool, which is useful for data-residency rules or vendor policies. We discuss the limits of that control in the risks section.

How Fugu benchmarks against the frontier

This is where honesty earns trust, because the marketing-friendly version of this story is misleading. Sakana's own careful wording is that Fugu Ultra stands "shoulder-to-shoulder" with Claude Fable 5 and Mythos Preview. It does not claim to beat them, and on two important tests it loses. Here are the published numbers.

Read it plainly. Fugu Ultra wins LiveCodeBench, Terminal Bench 2.1, GPQA-Diamond, and CharXiv Reasoning. Fable 5 leads on SWE-Bench Pro by a wide margin (86.0 versus 73.7) and on Humanity's Last Exam (53.3 versus 50.0). The honest summary is that Fugu matches the frontier without surpassing it, and the gap on real-world software engineering tasks (SWE-Bench Pro) is the one buyers should weigh most carefully. One caveat from independent reviewers: the Fable 5 and Mythos figures are provider-reported scores, not the result of a direct head-to-head test, so treat all cross-vendor benchmark tables with appropriate skepticism.

Beyond the standard suite, Fugu Ultra posted strong results elsewhere. On an AutoResearch benchmark it reported a best mean of 0.9774 bits-per-byte, ahead of three undisclosed frontier baselines. In a Rubik's Cube reasoning test, it solved the puzzle in 19 steps, the fewest among the four models tested (alongside Gemini 3.1 Pro, Opus 4.8, and GPT-5.5). The CharXiv figure of 86.1 in the table above is attributed to Mythos Preview rather than Fable 5, a small detail that illustrates how scattered the public numbers are.

Pricing and access

Fugu is sold both as a subscription and pay-as-you-go through the console at console.sakana.ai, using a single OpenAI-compatible API.

• Subscriptions: roughly $20/month (Standard), $100/month (Pro), and $200/month (Max).
• Pay-as-you-go, Fugu Ultra, per 1M tokens: $5 input / $30 output at standard context, rising to $10 input / $45 output for contexts of 272K tokens and above.
• Fugu (base) is significantly cheaper than Ultra.
• A launch promotion offered a free second month for subscribers who signed up before the end of July 2026.

Two things deserve a flag. First, the heaviest Ultra tasks can cost up to about $10 per message, because a single request can fan out into multiple model calls (more on that below). Some early users reported burning through the $20 tier in under five hours of intensive use. Second, the initial launch excluded EU users, with broader rollout to follow, and some users reported day-one errors on the site. If you are evaluating Fugu for production, confirm regional availability before you build around it.

For teams that prefer direct, predictable access to specific models rather than an orchestration subscription, the Truescho shop offers Claude Max, ChatGPT Plus, and Gemini subscriptions, which can be a cleaner fit for simple, high-volume work where orchestration overhead is wasted.

A practical business use case

Consider a mid-sized software company running automated code review on every pull request. With a single model, the team gets a fixed quality ceiling and a single dependency. One beta tester's experience captures the upside of orchestration here: "Where other tools flag about three issues, Fugu surfaced more than twenty." The verification stage, multiple models cross-checking each other, tends to catch more genuine problems than any one model reviewing alone.

The trade-off is cost and predictability. Code review on a busy repository is high-volume, and Fugu's fanout means each review may quietly spawn several specialist calls. For a small number of high-stakes, hard reviews (security-sensitive code, complex refactors), the extra spend buys real quality. For thousands of trivial diffs a day, a single mid-tier model is almost certainly the better economic choice. The smart pattern many teams will land on is selective orchestration: route the hard, expensive-to-get-wrong work to Fugu Ultra, and keep the routine high-volume traffic on a cheaper direct model.

Risks and honest criticisms

Fugu is genuinely clever, but it is not free of trade-offs, and the launch coverage that ignores them does buyers a disservice.

• Token fanout multiplies cost invisibly. A single Fugu Ultra request can consume 4 to 6 times the tokens of a direct single-model call, and may spawn 3 to 5 parallel specialist calls. Inside a flat subscription that cost is hidden until your quota disappears faster than expected.
• Margin stacking. Sakana pays full provider API rates and adds an orchestration margin on top. For simple tasks, that makes Fugu structurally more expensive than calling a model directly.
• A third-party quality ceiling. Fugu is only ever as good as the public models it can reach. It is not a sovereign capability; it is distributed dependence. If the best available models degrade or disappear, so does Fugu's edge.
• Opacity. You cannot see which underlying model produced any given output, and Sakana does not disclose the pool's composition (how much is closed versus open). The agent opt-out feature helps with compliance, but the lack of per-output transparency is a legitimate concern for regulated industries.
• Latency. Coordinating, verifying, and synthesizing across several models takes time. For interactive use the higher latency of Ultra is a real cost.

None of these are disqualifying, but together they define where Fugu fits: complex, high-value, vendor-risk-averse workloads, not high-volume commodity queries.

Who should try Fugu, and who should not

A simple decision frame. Fugu makes sense if your work is genuinely multi-step and high-stakes, if you value resilience against losing access to any single provider, and if you can tolerate variable per-request cost in exchange for higher answer quality. Fugu is wasteful if your workload is mostly simple, high-volume queries where a single cheaper model already does the job well; there, the fanout cost buys you little.

To understand the broader pattern Fugu represents, read our companion piece on AI orchestration models, and for a head-to-head with the frontier, see our Sakana Fugu Ultra vs Claude Fable 5 and Mythos comparison.

Frequently asked questions

What is Sakana Fugu and how does it work?

Fugu is a roughly 7-billion-parameter conductor model from Sakana AI. Rather than answering everything itself, it decides when to respond directly and when to assemble a team of expert models, coordinating them through a selection, delegation, verification, and synthesis loop, then returning one unified answer via an OpenAI-compatible API.

Is Fugu a real model or just a router?

It is a real, trained model, not a hand-coded router. Fugu is a compact language model trained with reinforcement learning to perform orchestration. The reward was whether the final answer was correct and well-formatted, so it learned its own routing and coordination strategies instead of following human-written if/else rules.

Who founded Sakana AI?

Sakana AI is a Tokyo-based company founded by David Ha (CEO) and Llion Jones (CTO). Both are former Google AI researchers, and Llion Jones is one of the co-authors of the 2017 paper "Attention Is All You Need," which introduced the Transformer architecture used across modern large language models.

What is the difference between Fugu and Fugu Ultra?

Fugu is the balanced default, tuned for strong performance at lower latency, ideal for coding, chatbots, and everyday tasks. Fugu Ultra targets maximum quality on hard, multi-step problems such as research reproduction and cybersecurity analysis, at higher latency and cost. You switch between them with a model-parameter change.

How much does Sakana Fugu cost?

Subscriptions run about $20, $100, and $200 per month. Pay-as-you-go for Fugu Ultra is roughly $5 input and $30 output per million tokens, rising to $10 and $45 for very large contexts. The base Fugu model is significantly cheaper, and the heaviest Ultra tasks can reach about $10 per message.

Does Fugu use GPT, Claude, and Gemini under the hood?

Fugu orchestrates a pool of publicly available expert models, and reference comparisons name GPT-5.5, Opus 4.8, and Gemini 3.1 Pro. The exact pool composition is undisclosed. Notably, Claude Fable 5 and Mythos are export-controlled and are not in the pool; Fugu matches their scores by coordinating other accessible models.

Can you tell which model answered your Fugu request?

No. Fugu deliberately presents itself as a single model, and you cannot see which underlying model produced a given output. Sakana also does not publish the pool's composition. This opacity is a recognized criticism, partly offset by the ability to exclude specific providers for compliance reasons.

Is Sakana Fugu available in the EU or my country?

At launch, Fugu excluded EU users, with broader availability promised later. Some users also reported day-one errors on the site. If regional access matters for your use case, verify availability directly in the console before building any workflow around it.

Conclusion

Sakana Fugu is one of the more interesting AI launches of 2026 precisely because it does not chase the usual prize. Instead of a bigger model, it offers a smarter conductor, a small trained orchestrator that matches frontier benchmarks by coordinating models you can actually access. The honest verdict is that it matches the frontier rather than beating it, that it loses to Fable 5 on real-world software engineering, and that its hidden fanout costs make it a poor fit for simple high-volume work. But for complex, high-stakes tasks, and for any organization worried about depending on a single AI provider, Fugu points to a genuinely useful new pattern. Whether the orchestrator becomes the new platform layer is the question every AI buyer and investor should now be asking.

Sources

• Sakana AI — Fugu release (official)
• The Decoder — Fugu orchestrates multiple LLMs to match Anthropic's Fable and Mythos
• GIGAZINE — Sakana Fugu multi-agent system
• BuildFastWithAI — Fugu review: the orchestration model that routes around export controls
• Krasa.ai — Fugu RL conductor (7B) orchestrates frontier models