Sakana Fugu Ultra: Japan's Multi-Agent System That Beats Claude Without Training a Giant Model

Tokyo didn't build a bigger brain. They built a smarter swarm.

On June 22, 2026, while the West was obsessing over parameter counts and $10B training runs, a small lab in Tokyo dropped what some engineers are calling "the most important AI release nobody noticed."

Sakana AI released Fugu and Fugu Ultra — not a foundation model, but an orchestration layer that sits in front of Claude Opus 4.8, GPT-5.5, Gemini 2.5 Pro, and others, and routes your query to the right specialist pool to produce frontier-grade output.

No trillion-parameter pretraining. No private cluster. Just pure orchestration.

And the numbers are uncomfortable: Fugu Ultra matches Claude Opus 4.8 on SWE-Bench Pro, competes with the rumored Claude Fable 5 system, and does it behind a single OpenAI-compatible API endpoint.

This is router pattern supremacy. This is Japan winning without playing America's game.

What is Sakana AI? The Tokyo Lab That Thinks Like Nature

Sakana AI was founded in 2023 by former Google researchers David Ha and Llion Jones (yes, the "Attention Is All You Need" Llion Jones). The name "Sakana" means fish in Japanese.

Their thesis from day one was heretical: AI shouldn't be built like a monolith. It should be built like nature.

Schooling fish. Ant colonies. Evolution. Collective intelligence emerging from simple agents following simple rules.

While OpenAI and Anthropic raced to build God in one model, Sakana was researching:

• Text-to-image evolution via evolutionary algorithms
• Model merging — breeding models together like organisms
• Continuous self-improving agents
• LLM-based scientific discovery loops

Sakana's previous work — AI Scientist, Darwin Godel Machine — already hinted at this philosophy. Fugu is the commercial culmination.

Headquartered in Tokyo with deep ties to the Japanese research ecosystem, Sakana has operated with unusual discipline: fewer press releases, more arXiv papers. That changed June 22nd.

What is Fugu? It's Not a Model. It's a Coordinator.

Let's kill the misconception immediately.

Fugu is not Sakana training a 500B model from scratch in Japan and naming it after a poisonous fish (though the metaphor of careful orchestration being potentially lethal if mishandled is apt).

Fugu is an orchestration engine.

Think of it like this: You have a pool of frontier models available via API. Each has strengths. Claude Opus 4.8 is absurdly good at code and reasoning. GPT-5.5 is strong at tool use and long context. Gemini 2.5 Pro is cheap, fast, and weirdly good at math.

Fugu's job is threefold:

1. Query Understanding & Decomposition When you send a prompt to Fugu's API, it doesn't just forward it. The router LLM parses intent, complexity, domain, and decomposes multi-step tasks into sub-tasks. A request like "Build me a RAG pipeline with evals" becomes: architecture planning (Claude), code generation (Claude), documentation (GPT), edge-case testing (Gemini pool).

2. Specialist Routing Each sub-task is dispatched to the model currently ranked highest for that capability in Sakana's live evaluation harness. This evaluation is continuous — Sakana published an open eval datastack on June 22, 2026, обновляемая daily. The pool adapts. If GPT-5.5 drops a new capability, Fugu's routing table updates automatically.

3. Synthesis & Verification Outputs from specialists are aggregated, cross-checked by a verifier model, and synthesized into a single coherent response. This is where Fugu Ultra diverges — it adds iterative debate and self-correction loops between agents, similar to Google's AlphaCode 2 but generalized.

The result: You call one endpoint. You get the collective intelligence of 5-8 frontier models without managing the chaos.

GitHub: SakanaAI/fugu — Fully open orchestration code, MIT licensed. Bring your own API keys.

Fugu vs Fugu Ultra: What's the Difference?

Sakana shipped two tiers, mirroring how OpenAI internally operates with Sol/Terra/Luna routing.

Fugu (Standard):

• Single-pass routing. Query -> Router -> Specialist -> Response
• Latency optimized. ~1.2-1.5x baseline latency of direct Claude call
• Cost optimized via cheaper models for easy sub-tasks
• Ideal for production APIs, chatbots, standard coding assistance

Fugu Ultra:

• Multi-agent debate, iterative refinement, self-consistency checks
• Can call models 3-8 times per user query internally
• Implements "Best-of-N sampling with verifier" loop
• Latency: 2-4x higher, cost: 2-3x higher, but reasoning ceiling massively lifted
• Designed to beat frontier evals: SWE-Bench Pro, GPQA Diamond, MATH frontier

If Fugu is a smart load balancer, Fugu Ultra is a full multi-agent parliament that argues until truth emerges.

This tiering is exactly what AI labs don't talk about. OpenAI's ChatGPT is not one model. It's a router deciding between GPT-5.5, o4, and smaller distilled models. Sakana just made the pattern explicit and open.

Benchmarks: Does It Actually Beat Claude?

Let's look at what Sakana released June 22, 2026. Note: Evaluations are from Sakana's continuously updated harness, not independent third-party yet. Treat with appropriate skepticism, but directionally interesting.

*Claude Fable 5 numbers are Sakana-estimated based on leaks / early API access. Not official Anthropic.

Three takeaways:

1. Fugu standard ~ matches Opus 4.8 at ~40% lower cost because it avoids using Opus for trivial sub-steps.
2. Fugu Ultra beats Opus 4.8 across the board and trades punches with Fable 5. Sakana's claim of beating Opus 4.8 on SWE-Bench Pro holds in their harness (73.4% vs 68.2%).
3. The router itself matters more than any single model. Swapping in better models in the pool automatically lifts Fugu's performance without retraining.

This is why Wall Street analysts called it "the most important AI release nobody noticed" — It undermines the economics of training ever-larger models. If you can get Fable-5-class performance by orchestrating Opus 4.8 + GPT-5.5 cleverly, why spend $500M training Fable 6 alone?

Code Example: Drop-In OpenAI Compatible

This is where it gets dangerous for incumbents. Fugu speaks OpenAI.

You don't need to rewrite your stack.

javascriptCopy
1// Before: Direct Claude via Anthropic SDK
2// After: Same interface, swarm intelligence underneath
3
4import OpenAI from 'openai';
5
6const client = new OpenAI({
7  baseURL: 'https://api.sakana.ai/v1/fugu', // or /v1/fugu-ultra
8  apiKey: process.env.SAKANA_API_KEY, // bring your Anthropic + OpenAI keys in dashboard
9});
10
11const response = await client.chat.completions.create({
12  model: 'fugu-ultra', // or 'fugu'
13  messages: [
14    { role: 'system', content: 'You are a senior distributed systems engineer.' },
15    { role: 'user', content: 'Design a fault-tolerant job queue in Go with exactly-once semantics using Redis Streams. Include failure modes.' }
16  ],
17  temperature: 0.2,
18});
19
20console.log(response.choices[0].message.content);
21// Behind the scenes: router -> Claude for architecture -> GPT-5.5 for Go idioms -> verifier loop -> response

Self-hosted? Also possible:

bashCopy
1git clone https://github.com/SakanaAI/fugu
2cd fugu
3pip install -e .
4export ANTHROPIC_API_KEY=...
5export OPENAI_API_KEY=...
6export GOOGLE_API_KEY=...
7
8python -m fugu.serve --model-pool config/ultra_pool.yaml --port 8000

The open-source repo includes the router prompts, the evaluation framework, and the synthesis logic. You plug in your own frontier keys. Sakana doesn't proxy your tokens if you self-host — the hosted version is for convenience.

Use Cases: Where This Actually Wins

Fugu isn't for everything. But for specific workloads, it's architecturally superior.

1. Complex Software Engineering SWE-Bench Pro is the tell. Real GitHub issues requiring repo-wide reasoning. Fugu Ultra's multi-agent debate catches hallucinations where a single model would ship broken code. This is a legit Copilot alternative for senior-level tasks.

2. Cost-Arbitraged Production Apps If you're burning $20k/month on Opus 4.8 for a customer support bot, Fugu standard can cut that to $8-10k by routing 70% of queries to cheaper models without quality drop. The router learns.

3. Research & Agentic Workflows Sakana's continuous evaluation harness means Fugu is ideal for research scaffolding. You can A/B test models per task slice and see live rankings. For teams building agents, it's a meta-agent layer.

4. Japanese / Multilingual Enterprise Sakana's Tokyo roots show — Fugu has strong Japanese language routing and has been benchmarked on Japanese legal/medical evals where GPT-5.5 alone struggles. If you're operating in APAC, this matters.

Limitations: The Poison in the Pufferfish

Fugu is named after the most dangerous delicacy in Japan for a reason. Mishandle it, you die (metaphorically, via latency bills).

1. Latency Variance Fugu Ultra can take 20-45 seconds for complex reasoning. Not suitable for real-time chat without streaming synthesis. Standard Fugu is better (~2-3s), but still slower than direct call.

2. API Key Sprawl & Cost Opacity You need keys for every model in the pool. Hosted version abstracts this but adds margin. Self-hosted means you manage 3-4 bills. And Fugu Ultra's internal calls can spike costs unpredictably — a "simple" prompt that triggers 6 internal calls costs 6x.

3. Non-Determinism Orchestration = variance. Same prompt routed twice might hit different specialists based on load or updated rankings. Harder to eval than single model. Sakana provides seed control, but it's not perfect.

4. Dependency on Frontier Labs Fugu's performance is derivative. If Anthropic deprecates Opus 4.8 or changes pricing, Fugu's benchmarks collapse until retuned. It's a meta-system, not a foundation. Sakana is betting Tokyo can stay ahead on routing while frontier labs do the expensive pretraining.

This is the core critique: Fugu is brilliant arbitrage, not fundamental capability creation. But as one engineer tweeted: "All of software is arbitrage done well."

Why Orchestration Is The Future, Not Bigger Models

We have hit the wall everyone whispered about in late 2025.

Scaling laws aren't dead, but marginal returns are brutal. Going from 70% to 74% on SWE-Bench Pro via pretraining might cost $300M. Getting there via better orchestration of existing models costs ~$0 in training compute.

Sakana, OpenAI (Sol/Terra/Luna), and recent DeepMind papers on "Mixture-of-Agents" are converging on the same insight:

The biggest unlock is no longer training a bigger model. It's routing, verifying, and synthesizing the models we already have.

It's cheaper, it's composable, and it's defensible via eval harnesses, not GPU hoards.

Japan, with limited access to massive compute clusters compared to US hyperscalers, had no choice but to innovate on architecture. Constraint bred creativity. Fugu is the result.

Expect every lab to ship their own Fugu in next 6 months. Sakana just got there first and open-sourced it.

FAQ

Q1: Is Fugu Ultra actually better than Claude Fable 5? In Sakana's eval harness, Fugu Ultra trades blows — it beats Fable 5 estimates on GPQA Diamond, loses narrowly on SWE-Bench Pro. But Fable 5 is not officially released, so these are leaked estimates. The honest answer: Fugu Ultra gets you ~90-95% of rumored Fable 5 performance today using models you already pay for. That's insane value.

Q2: Do I need to give Sakana my Anthropic / OpenAI API keys? For hosted api.sakana.ai, you can either bring your own keys (BYOK) stored encrypted, or pay Sakana's bundled rate where they proxy. For self-hosted open-source SakanaAI/fugu, you run the router locally and keys never leave your VPC. Self-host is recommended for enterprise.

Q3: How is Fugu different from LangGraph or other orchestration frameworks? LangGraph is a framework to build multi-agent workflows. You write the graph. Fugu is a pretrained orchestration system with learned routing policies, verification loops, and a continuously updated evaluation that decides which model handles which sub-task. You don't build the graph — Sakana already did. Think framework vs product.

Q4: Can Fugu orchestrate open-source models like Llama 4 or Qwen3? Yes. The pool is configurable YAML. Sakana's default pool uses frontier closed models because they win benchmarks, but community forks already run Fugu with Llama 4 400B and Qwen3 235B as specialists for cost-sensitive deployments. Performance drops ~5-8% vs frontier pool, but cost drops 80%.

Final Verdict: Watch Tokyo

I called this the most important release of June 2026 because it reframes the game.

For 3 years, the narrative was: bigger cluster wins. US hyperscalers win. Japan and Europe can't compete.

Sakana just proved you can compete by being smarter, not richer. Multi-agent orchestration, continuous evaluation, and ruthless routing pragmatism beat monolithic scaling.

Fugu will not replace Claude Opus 4.8 or GPT-5.5. It makes them both more valuable by using them correctly. The future is not one model to rule them all — it's a coordinated swarm of specialists behind one boring OpenAI-compatible API.

Poisonous if mishandled. Delicious if mastered.

Just like fugu.

---

Want to ship AI systems that actually understand orchestration patterns like this — not just call one model and pray? I break down frontier architectures every week.

Check the toolkit: essamamdani.com/tools — production-grade prompts, routers, and eval harnesses I've battle-tested building AI products in 2026.