Japan Just Built an AI Monster That Rivals Claude: The Rise of Sakana Fugu

By — IndianAI.in Editorial Team
Published: June 27, 2026

On June 22, 2026, the Tokyo-based artificial intelligence lab Sakana AI did something the world was not entirely prepared for. They unveiled Fugu — not merely another large language model, but an entirely new breed of AI architecture that learns how to orchestrate intelligence rather than simply generate it. The benchmarks were startling. The implications, even more so.

Named after the Japanese pufferfish — a delicacy that can kill if prepared incorrectly, but is exquisite when handled by a master — Fugu represents a philosophical and technical departure from everything the AI industry has taken for granted over the past five years. And in doing so, it has forced a fundamental question: Is the era of the monolithic model coming to an end?

This article is a deep, humanistic, research-driven exploration of what Sakana Fugu is, how it works, why it matters, and what it means for India's AI ecosystem — written without hype, without fear, and with the clarity that only genuine understanding provides.

The Core Idea: What Is Sakana Fugu?

Let us begin with the most important distinction — because nearly every headline in the West got this wrong.

Fugu is not a large language model in the traditional sense.

It is not a single neural network trained on trillions of tokens, distilled into a dense transformer, and served as a monolithic endpoint. That is Claude. That is GPT. That is Gemini.

Fugu is something different. It is a learned multi-agent orchestration system — a "conductor" model trained not to answer questions directly, but to decide which model should answer, how to decompose the question, when to verify the output, and how to synthesise results from multiple sources into a coherent final answer.

David Ha, CEO and co-founder of Sakana AI — formerly of Google Brain and a long-time advocate of neuroevolution and alternative AI architectures — described it plainly in the launch documentation:

"Fugu is essentially an LLM trained to engage various LLMs within an agent pool, including recursive instances of itself."

In simpler terms: imagine you walk into a room filled with the world's greatest specialists — a surgeon, a physicist, a poet, a coder, a philosopher. You ask a question. One of them alone might give you an incomplete answer. But a master coordinator — someone who knows exactly whom to ask, in what order, how to verify each partial answer, and how to weave it all together — will give you something far greater than any individual specialist could.

Fugu is that master coordinator. Trained end-to-end to orchestrate intelligence.

Architecture: How Fugu Actually Works

The Orchestrator Model

At the heart of Fugu is a relatively compact conductor model that has been trained to:

1. Analyse an incoming prompt and determine its nature — is this a coding task? A research question? A creative brief? A multi-step reasoning problem?
2. Decompose the task into sub-tasks that can be routed to different specialist models.
3. Route each sub-task to an appropriate model from a dynamic pool of frontier LLMs.
4. Verify outputs — sometimes through a secondary model, sometimes through self-consistency checks.
5. Synthesise the results into a single, coherent, high-quality response.

The genius of this approach is that the conductor model itself is learned. It is not a hand-coded rules engine. It improves over time, it adapts to new models in its pool, and it learns from its own mistakes.

The Model Pool

Fugu's current agent pool includes some of the most capable models in existence:

• GPT-5.5 (OpenAI)
• Claude Opus 4.8 (Anthropic)
• Gemini 3.1 Pro (Google DeepMind)
• And several specialised open-weight models for specific domains

Notably absent from the pool are Claude Fable 5 and Claude Mythos 5 — Anthropic's most powerful models. This is not a technical limitation. On June 12, 2026, a US government export control directive suspended access to both models, effectively taking them offline just three days after their launch. Sakana cannot orchestrate what it cannot reach.

This geopolitical twist, which we will explore later, has turned Fugu from an interesting research project into a genuinely strategic alternative.

Two Flavours: Fugu and Fugu Ultra

Fugu is designed for latency-sensitive use cases — interactive chat, real-time coding assistance, quick analysis. Fugu Ultra is built for the hard stuff: multi-hour research workflows, scientific reasoning, agentic coding, and tasks where accuracy matters more than speed.

The Benchmark Story: Where Fugu Wins, Where It Doesn't

The Numbers

Sakana published a comprehensive benchmark suite. Here is the human-readable truth of what those numbers mean.

Fugu Ultra vs. The Field

What These Numbers Actually Mean

The wins are real. On SWE-Bench Pro — one of the most demanding software engineering benchmarks available — Fugu Ultra scores 73.7%, beating every single model in its own pool. On LiveCodeBench, it hits 93.2%, outperforming GPT-5.5 by nearly 8 points and Gemini 3.1 Pro by nearly 5 points. On GPQA-Diamond (a test of 198 graduate-level multiple-choice questions in biology, physics, and chemistry), it achieves 95.5%, surpassing even the formidable Claude Opus 4.8.

These are not marginal improvements. These are structural advantages. Orchestration, when done well, produces emergent intelligence — a whole that is genuinely greater than the sum of its parts.

But — and this is crucial — the wins are not universal.

On cybersecurity benchmarks (CTI-REALM), Claude Opus 4.8 still edges ahead (69.6 vs. 69.4). On long-context recall (MRCRv2), GPT-5.5 leads (94.8 vs. 93.6). And on the hardest general knowledge benchmark — Humanity's Last Exam — Fugu Ultra (50.0) essentially ties with Opus 4.8 (49.8) but falls short of Claude Fable 5 (53.3).

The Mythos 5 Caveat

Here is where the comparison becomes complex — and where intellectual honesty is essential.

Sakana's published benchmarks compare Fugu Ultra against Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro. They do not compare against Claude Mythos 5 — because Mythos 5 is not in Fugu's model pool and is currently unavailable due to US export controls.

However, independent data from BenchLM.ai (which ranks Mythos 5 first out of 124 models evaluated) shows:

The honest assessment: On raw reasoning power, Claude Mythos 5 — if it were available — would likely outperform Fugu Ultra on the most demanding benchmarks. The gaps are meaningful: 6.6 points on SWE-Bench Pro, 5.9 on Terminal-Bench, 9 full points on Humanity's Last Exam.

But Mythos 5 is not available. It has been offline since June 12, 2026, with no resumption date announced. Fugu Ultra is available now.

This transforms the comparison from a technical one to a strategic one. The question is no longer "Which model is more powerful?" but rather "Which model can you actually use?"

The Geopolitical Context: Why Fugu Matters Right Now

On June 12, 2026, the United States government issued an export control directive that suspended access to Claude Fable 5 and Claude Mythos 5 — Anthropic's most advanced models — for users outside specific approved jurisdictions.

The rationale, as publicly stated, was related to national security concerns around the proliferation of frontier AI capabilities. The practical effect was immediate and dramatic: developers, enterprises, and researchers around the world — including across India, Southeast Asia, Africa, the Middle East, and Latin America — lost access to models they had already begun integrating into critical workflows.

This is where Fugu enters the story not just as a technology, but as a statement of strategic autonomy.

Sakana AI is a Japanese company. Japan, like India, occupies a complex position in the global AI landscape — technologically advanced, deeply integrated with Western AI ecosystems, but increasingly aware of the risks of dependency on any single nation's infrastructure.

David Ha and his team built Fugu with a philosophy that now looks prescient: orchestration over reliance. Instead of betting everything on one model from one provider in one country, Fugu creates a layer of abstraction that can dynamically route around availability gaps, geopolitical disruptions, and regulatory changes.

This is not anti-American. It is not anti-anyone. It is pro-resilience. And for a country like India — which is racing to build its own AI infrastructure, train its own models, and serve its billion-plus population — that philosophy resonates deeply.

Pricing: A Cost Story That Matters

Fugu is priced at $5.00 input / $30.00 output per million tokens — significantly cheaper than comparable frontier models. There is a long-context surcharge (2× input / 1.5× output for prompts over 272K tokens), but for standard workloads, the pricing is competitive.

For Indian startups, research institutions, and enterprises operating on tighter margins than their Silicon Valley counterparts, this pricing differential is not trivial. It is the difference between a prototype that scales and one that breaks the budget.

What the Experts Are Saying

The Technical Verdict

Independent reviewers have been careful to separate Fugu's genuine achievements from its marketing.

Bind AI's analysis concluded:

"For developers who need strong results today, Fugu Ultra is the most capable option currently available. For teams willing to wait on Mythos 5 resumption and who need verified scores before committing, patience is justified. The better model, by the numbers, is Mythos 5. The only available model, by circumstance, is Fugu."

Coursiv's review offered a balanced perspective:

"Fugu may be worth evaluating if your team needs a multi-model AI layer for coding, research, or difficult analysis. It should not be treated as a one-to-one substitute for Claude Fable 5. The main reason to test Fugu Ultra is to see whether learned orchestration changes outcomes on your own workload while reducing dependence on a single model provider."

VentureBeat framed it within the broader industry context:

"An orchestration system is ultimately limited by the inherent capabilities of the models in its pool — a reality reflected in Sakana's own benchmark evaluations against standalone frontier models."

The Honest Quirk

One fascinating detail from the benchmarks: on certain tasks like SciCode, the standard Fugu (not Ultra) actually scored higher than Fugu Ultra. This suggests that more orchestration is not always better — and that Sakana's own system is still learning when to apply its full arsenal and when a lighter touch suffices.

This is not a weakness. It is a sign of a system that is genuinely adaptive rather than brute-force.

What This Means for India

A Model for Model Independence

India's AI strategy — articulated through initiatives like IndiaAI Mission, the National AI Portal, and the push for sovereign AI infrastructure — has always grappled with a central tension: how to stay globally competitive while building locally relevant capabilities.

Fugu's architecture offers a template that Indian AI labs could adapt and build upon:

• Multi-provider orchestration reduces dependency on any single Western AI company.
• Learned routing means the system improves with use, adapting to Indian languages, cultural contexts, and domain-specific needs.
• Open-compatible API (OpenAI-compatible endpoint) means existing Indian AI applications can integrate without rewriting their codebase.
• Cost efficiency at $5/$30 per million tokens makes frontier-grade AI more accessible for Indian developers.

The Indian Developer's Dilemma

For years, Indian AI developers have faced a frustrating choice:

• Use frontier Western models — expensive, sometimes geopolitically constrained, and rarely optimised for Indian languages or contexts.
• Build from scratch — capital-intensive, time-consuming, and risky in a fast-moving landscape.
• Use open-weight models — capable but often lagging behind frontier capabilities.

Fugu represents a fourth path: leverage the best of what exists, but through an intelligent layer that you control. This is not a replacement for building Indian foundational models — it is a complement. A bridge. A way to compete today while building indigenous capacity for tomorrow.

Japanese-Indian AI Collaboration

The parallels between Japan's AI journey and India's are striking:

Sakana AI's success demonstrates that strategic autonomy in AI is achievable without isolation. Japan did not cut itself off from Western AI research. It did the opposite — it integrated deeply, but built a layer of orchestration that gives it control.

India can — and should — do the same.

The Bigger Picture: What Fugu Teaches Us About AI's Future

Lesson 1: Intelligence Is a Coordination Problem

For too long, the AI industry has been obsessed with scale — bigger models, more parameters, more compute, more data. Fugu suggests that coordination may be as important as scale. The ability to intelligently route, verify, and synthesise across multiple models can produce results that exceed any single model in the pool.

This is not a new idea in computer science — distributed systems, ensemble methods, and mixture-of-experts have been studied for decades. But Fugu is one of the first production-grade systems to make this work at frontier level with learned orchestration.

Lesson 2: Geopolitics Will Shape AI Architecture

The suspension of Claude Mythos 5 was a shock to many in the industry. It should not have been. As AI becomes strategically critical — on par with semiconductors, energy, and defence — every nation that relies on imported AI infrastructure is vulnerable to supply-side disruptions.

Fugu's architecture is a hedge against that vulnerability. By abstracting away individual model providers, it creates a buffer between the user and geopolitical turbulence.

Lesson 3: The Best Model Is the One You Can Actually Use

This sounds trite. It is not.

The AI community has a tendency to obsess over leaderboard positions and benchmark scores, as if raw capability were the only metric that matters. But availability, reliability, cost, latency, regulatory compliance, and geographic accessibility are equally important.

Fugu Ultra may not be the most powerful AI system ever built. But it is available, it is affordable, and it works. For millions of developers and enterprises — especially in the Global South — that combination matters more than a 6-point lead on a benchmark they will never run.

Where Fugu Excels: Practical Use Cases

1. Complex Software Engineering

On SWE-Bench Pro — a benchmark that evaluates an AI's ability to resolve real-world GitHub issues — Fugu Ultra's 73.7% outperforms every model in its pool. For Indian SaaS startups and product teams, this translates to:

• Faster bug resolution
• More reliable code reviews
• Better handling of multi-file, multi-step coding tasks

2. Scientific Reasoning and Research

On GPQA-Diamond (graduate-level science questions), Fugu Ultra's 95.5% is best-in-class among accessible models. Indian research institutions — from IITs to IISc to private R&D labs — can leverage this for:

• Literature review and synthesis
• Hypothesis generation
• Experimental design assistance

3. Mathematical Problem Solving

With 99.0% on MATH-500 and 90.0% on AIME 2025, Fugu Ultra demonstrates near-perfect mathematical reasoning. For India's vast ecosystem of EdTech platforms, competitive exam preparation, and STEM education, this is directly applicable.

4. Multi-Agent Agentic Workflows

Because Fugu is itself an orchestration system, it is naturally suited for agentic workflows — tasks that require planning, tool use, memory, and multi-step execution. On TerminalBench 2.1, Fugu Ultra's 82.1% leads all accessible models.

Limitations and Honest Concerns

No technology is without its caveats, and Fugu has several that deserve honest discussion.

1. Not a Single Model

Fugu's strength — orchestration — is also its complexity. When a system routes your prompt through multiple models, you lose some degree of predictability. The same prompt may be handled differently depending on routing decisions, model availability, and latency conditions. For applications that require deterministic behaviour, this is a challenge.

2. Geographic Restrictions

Fugu is currently unavailable in the EU and EEA due to ongoing GDPR compliance work. For Indian companies operating in European markets, this creates a compliance gap that needs attention.

3. Vendor-Published Benchmarks

All of Fugu's benchmark results, while impressive, are vendor-published. Independent third-party verification is still forthcoming. The AI community should treat these numbers as indicative but not definitive until replicated by neutral evaluators.

4. Pool Dependency

Fugu is only as good as the models in its pool. If the most capable models become unavailable (as happened with Mythos 5), Fugu's ceiling is lowered. Long-term, Sakana will need to ensure a diverse, resilient, and continuously updated model pool.

5. No Open Weights

Unlike some of Sakana's earlier research (which focused on open-weight, evolutionary approaches), Fugu is a closed, API-only product. For the open-source AI community — particularly strong in India — this may limit adoption.

The Road Ahead: What Comes After Fugu

Sakana AI has already signalled that Fugu is just the beginning. The company's long-term vision includes:

• Self-improving orchestration: Systems that learn from production usage and automatically refine their routing strategies.
• Expanded model pools: Integration with more frontier and specialised models as they become available.
• On-premise deployments: For enterprises with strict data sovereignty requirements.
• Regional optimisations: Including models fine-tuned for Indian languages, Japanese, and other non-English languages.

For India, the opportunity is clear. Sakana's approach — intelligent orchestration over monolithic dependence — aligns perfectly with the strategic goals of IndiaAI Mission, the push for indigenous AI capability, and the need for cost-effective, accessible frontier AI.

Final Verdict: A Genuine Contender, Not a Universal King

Let us be clear.

Fugu is not the most powerful AI system ever built. Claude Mythos 5, were it available, would likely outperform it on the hardest benchmarks. GPT-5.5 has advantages in specific domains. Claude Opus 4.8 holds its own in cybersecurity and long-context tasks.

But Fugu is the most strategically important AI release of 2026.

It demonstrates that intelligence can be coordinated as effectively as it can be generated. It proves that a well-designed orchestration layer can compete with — and sometimes surpass — the world's most expensive monolithic models. And it arrives at a moment when the AI world is crying out for alternatives to American-dominated infrastructure.

For India, Fugu is both an inspiration and a blueprint. An inspiration because a Japanese startup, operating from Tokyo, with a fraction of the resources of OpenAI or Anthropic, has built something genuinely world-class. A blueprint because the architecture — learned orchestration over a diverse model pool — is something Indian AI labs can study, adapt, and perhaps one day surpass.

The pufferfish, in Japanese culinary tradition, is a dish that demands respect. It can kill. It can delight. But only a master chef knows how to prepare it.

Sakana AI has cooked something remarkable. Whether it becomes a global staple or a regional delicacy depends on what the rest of the world — including India — does next.

References

1. Sakana AI, "Fugu: A Learned Multi-Agent Orchestration System", June 22, 2026 — sakana.ai/fugu
2. VentureBeat, "No Claude Fable 5? No problem: Sakana achieves frontier performance with new Fugu multi-model, auto synthesis system", June 22, 2026
3. Coursiv, "Sakana AI Fugu Review: Fugu Ultra vs Claude Fable 5", June 22, 2026
4. Bind AI, "Sakana Fugu vs Claude Mythos: Which Is the Better Model?", June 23, 2026
5. Labellerr, "Sakana Fugu: The Model That Outsmarts Claude Fable 5", June 23, 2026
6. BenchLM.ai, "Global AI Model Rankings", June 2026
7. Anthropic, "Claude Mythos 5 and Fable 5: Technical Report", June 2026
8. US Department of Commerce, "Export Control Directive on Frontier AI Models", June 12, 2026
9. IndiaAI Mission, "National AI Strategy Update", Ministry of Electronics and Information Technology, 2026
10. David Ha, "Evolutionary Approaches to AI Architecture", Google Brain / Sakana AI (Various Publications)

This article is published by IndianAI.in — India's leading platform for independent, thoughtful, and human-centric analysis of artificial intelligence. We believe in AI that serves people, not the other way around.

📩 Have thoughts on this? Write to us or join the conversation on our community forum.