What Is Sakana Fugu? Multi-Agent AI Model Explained

Key Takeaways

• Sakana Fugu is not a standalone LLM. It is a learned orchestration system that coordinates multiple frontier AI models behind one single API.
• Two variants exist: Fugu (balanced, low latency) and Fugu Ultra (max performance, deeper agent pool).
• Fugu Ultra scores 73.7 on SWE-Bench Pro, beating GPT-5.5 (58.6) and Claude Opus 4.8 (69.2) on Sakana’s own benchmarks.
• Pricing starts at $20/month (subscription) or $5 input / $30 output per 1M tokens (pay-as-you-go for Fugu Ultra).
• Access is via a fully OpenAI-compatible API at https://api.sakana.ai/v1. No SDK migration required.
• Sakana AI is a Tokyo-based lab founded by former Google Brain researchers David Ha and Llion Jones.

What Is Sakana Fugu?

Sakana Fugu is a multi-agent orchestration model from Sakana AI, a Tokyo-based AI research lab.

It launched on June 22, 2026, and it works differently from every AI model you have used before.

Instead of running a single large language model, Fugu routes your request to a pool of specialized frontier models. It coordinates their outputs and returns one unified response.

From your side, you call one API endpoint. Everything else happens behind the scenes.

“From the outside, you simply call one model. On the inside, a coordinated system of experts is doing the work.” — Sakana AI, June 2026 release announcement

Fugu is not a frontier model. It does not replace Opus 4.8, GPT-5.5, or Gemini. It coordinates them. Fugu is the conductor. Those models are the orchestra.

Who Is Sakana AI?

Sakana AI was founded by David Ha and Llion Jones, both former Google Brain researchers.

The lab is headquartered in Tokyo and focuses on nature-inspired AI research. Fugu is its first major commercial product.

The name “Fugu” refers to the Japanese pufferfish, known for its complexity and the precision required to handle it safely. The metaphor fits the product.

How Fugu Differs from a Standard LLM

Most AI models work in a straight line. You send a prompt, one model processes it, you get a reply.

Sakana Fugu works differently. Here is the core distinction at a glance:

The key insight: the complexity of a multi-agent system never reaches your code. You write one API call. Fugu handles the rest.

How Does Sakana Fugu Work?

Fugu is itself a language model. But it is trained to call other LLMs, not just answer you directly.

Here is exactly what happens when you send a request:

1. Your request arrives at the Fugu endpoint.
2. Fugu evaluates whether to answer directly or assemble agents.
3. If agents are needed, Fugu selects models from its pool and assigns roles.
4. Each agent works on its assigned sub-task simultaneously.
5. Fugu synthesizes the outputs into one coherent response.
6. You receive one reply with no visibility into which models ran.

The routing logic is proprietary. Sakana does not expose which model handled which part of your request on a per-query basis.

TRINITY: The Role Assignment System

TRINITY is a lightweight coordinator, roughly 0.6B parameters, evolved using CMA-ES optimization.

It assigns dynamic roles to agents across multiple turns:

• Thinker: reasons through the problem
• Worker: executes the task
• Verifier: checks the output for accuracy

TRINITY adapts role assignments based on what the task actually requires. It covers coding, math, reasoning, and knowledge tasks. It was published at ICLR 2026.

Conductor: The Coordination Learner

Conductor is a 7B model trained with reinforcement learning.

It discovers natural-language coordination strategies entirely on its own, without hand-coded rules. Two key capabilities set it apart:

• It builds focused, task-specific prompts for each model in the pool.
• It can call itself recursively to scale compute at inference time.

Together, TRINITY and Conductor make Fugu’s orchestration learned, not scripted. This is the core research claim Sakana makes for the system.

Fugu vs. Fugu Ultra: Which One Should You Use?

Sakana ships two variants under the same OpenAI-compatible API. The choice comes down to one question: do you need speed or quality?

Use Fugu when your workload is latency-sensitive or when your team needs agent opt-out for compliance or data privacy.

Use Fugu Ultra when you are running hard research tasks, multi-step agentic workflows, or pipelines where answer depth matters more than response time.

Fugu Ultra v1.1 Update

On July 24, 2026, Sakana released Fugu Ultra v1.1.

It claims up to 7.9 benchmark point improvements over v1.0, with the strongest gains on ProgramBench and TerminalBench 2.1. Pricing is unchanged. The original snapshot (fugu-ultra-20260615) is now formally versioned as v1.0.

Sakana Fugu Benchmark Results

Fugu Ultra leads 10 of 11 published benchmarks against its own agent pool members.

Important caveat: all numbers below are Sakana-reported as of June 2026. Independent third-party replication is not yet available. Do not make production routing decisions from this table alone. Run your own evaluation on 20 to 50 representative tasks first.

Core Benchmark Scores

What the Benchmark Table Does Not Show

Three caveats matter before you read these numbers as definitive:

• Fable 5 is absent from the comparison. Claude’s Fable 5 model scores 86.0 on SWE-Bench Pro, nearly 13 points above Fugu Ultra’s 73.7. Fable 5 is excluded from Fugu’s agent pool due to US export control restrictions, not because Fugu outperforms it.
• More orchestration is not always better. On SciCode and a few other benchmarks, the standard Fugu scores higher than Fugu Ultra. “More agents” does not equal “better output” in every case.
• Scores are system-level, not model-level. A high benchmark number reflects Fugu’s routing, delegation, and synthesis quality, not any single underlying model’s raw ability.

Notable Demo Results

These are Sakana-reported results from launch demonstrations. Treat them as directional evidence, not production guarantees.

• Fugu outperformed three frontier models and a 2100-Elo Stockfish engine at blindfold chess.
• In a time-series trading simulation, Fugu grew $10,000 to $11,943 across 50 weeks, a 19.43% return that beat every model in the comparison.

Why Sakana Built Fugu: The Export Control Context

The launch timing was not accidental.

Ten days before Fugu’s June 22, 2026 release, the US government ordered Anthropic to suspend access to Fable 5 and Mythos Preview for non-US entities.

Sakana AI, based in Tokyo, positioned Fugu as a direct response. The core argument: frontier capability without the risk of export controls or single-vendor dependency.

The agent pool is designed to be swappable. Newer models fold in over time. Providers who restrict access can be routed around.

This is a genuine architectural advantage. But it carries one honest limitation:

Fugu still rents its intelligence from the same underlying vendors. A pool that loses access to multiple providers simultaneously would lose significant capability. Resilience comes from pool diversity, not from the orchestration layer itself.

Sakana Fugu Pricing

Fugu offers two pricing structures. Both include access to Fugu and Fugu Ultra.

Subscription Plans

Subscription plans are best for regular daily or professional use where a predictable monthly cost matters more than exact token counting.

Subscribe before July 31, 2026, to receive a free second month at your initial tier.

Pay-As-You-Go API Pricing (Fugu Ultra)

Pay-as-you-go targets production workloads where token volume is easier to predict than a flat monthly fee.

For the standard Fugu model on pay-as-you-go, you are charged the rate of the top-tier model involved in your request. Sakana does not stack fees when multiple agents run on the same task.

Is Sakana Fugu Free?

There is no permanent free tier. The only free offer at launch is the second-month subscription bonus for users who sign up before July 31, 2026.

How to Access Sakana Fugu: API Setup

Fugu uses a fully OpenAI-compatible API. You do not need to migrate any SDK or rewrite existing client code.

Step-by-Step API Setup

Step 1: Create your account

Sign up at console.sakana.ai. Generate your API key from the account dashboard.

Step 2: Set the base URL

Point your existing OpenAI client at the Fugu endpoint:

Base URL: https://api.sakana.ai/v1

Step 3: Choose your model ID

Use one of the following model strings depending on which variant you want:

fugu                      # Standard Fugu (balanced, lower latency)
fugu-ultra-20260615       # Fugu Ultra v1.0

Step 4: Swap your API key

Replace your existing provider API key with your Sakana key. No other code changes are required. Your existing OpenAI-format calls work as-is.

Compatible Tools and Integrations

Because Fugu uses an OpenAI-compatible API, it works out of the box with the following tools:

• OpenCode — Sakana’s recommended coding integration
• Codex CLI — works without any configuration changes
• Cursor — set Fugu as your model provider in settings
• VS Code AI extensions — any extension that accepts a custom base URL
• Any existing client or coding harness built for OpenAI

Agent Opt-Out for Compliance Teams

Standard Fugu lets you exclude specific providers from its agent pool.

This is useful for teams with data residency, privacy, or regulatory requirements that restrict which AI providers can process your data.

Agent opt-out is not available on Fugu Ultra. Its pool is fixed and cannot be modified.

Is Sakana Fugu Open Source?

Fugu is not open source as a commercial product. The API is closed.

However, Sakana has made the underlying research publicly available:

• Technical report — arXiv 2606.21228
• TRINITY paper — published at ICLR 2026
• Conductor paper — published at ICLR 2026

The research that powers Fugu is peer-reviewed and readable by anyone. The orchestration system built on top of that research is proprietary.

Fugu is not available on HuggingFace as a downloadable model weight. It runs exclusively as a hosted API through Sakana’s platform. There is no local or self-hosted version available at launch.

Real-World Use Cases

Where Standard Fugu Performs Best

Standard Fugu is the right choice for latency-sensitive or compliance-constrained workloads:

• Everyday coding and code review inside tools like Codex or Cursor
• Chatbot backends where response speed is critical
• Teams that need provider opt-out to meet data residency requirements
• Interactive coding assistants with high request volume

Where Fugu Ultra Performs Best

Fugu Ultra is designed for tasks where depth and accuracy matter more than speed:

• AI research and scientific paper reproduction
• Complex multi-step reasoning and analysis pipelines
• Production workflows where answer quality is the primary success metric
• Tasks that benefit from multiple expert agents verifying each other’s output

Sakana Fugu vs. Competitors

Fable 5 scores 86.0 on SWE-Bench Pro. Fugu Ultra scores 73.7. Fable 5 is not in Fugu’s pool. The two systems are not in a direct head-to-head relationship.

Limitations to Know Before Committing

Latency is meaningfully higher. Orchestrating multiple agents takes more time than a single model call. If you need sub-second response times, Fugu is not the right fit today.

Routing is opaque by design. You cannot see which model handled which part of your request. Teams that need full query-level explainability will find this constraining.

Benchmarks are vendor-reported. No independent third-party lab has replicated Sakana’s June 2026 numbers at the time this article was written. The scores are evidence to test against your own workload, not finished proof.

The pool still depends on underlying vendors. Fugu reduces single-vendor dependency. It does not eliminate it. If access to multiple providers in the pool is restricted simultaneously, capability drops.

Fugu Ultra has no compliance opt-out. If your legal or data team requires specific provider exclusions, Fugu Ultra cannot accommodate that. Use standard Fugu instead.

Working Through the Model Decision With a Partner

Evaluating Sakana Fugu alongside Claude, GPT-5.5, and other frontier models is not a one-afternoon task. Benchmark tables tell you what a model scored on a curated test. They do not tell you how it performs on your specific workload, at your token volumes, with your data.

Phos AI Labs is an embedded AI consulting partner for mid-market companies. We are CCA-F certified by Anthropic, members of the Anthropic Claude Partner Network, and one of the first ten firms globally in the OpenAI Select Partner Network. We evaluate AI tools against real workloads, not marketing benchmarks, then build the foundations your team needs to use them well.

• Model evaluation: We test your representative tasks against Fugu Ultra, Claude Opus 5, and others before recommending anything
• AI strategy: We build the decision framework before any tool is deployed
• Team training: We train your team on the tools that actually fit your workflows
• AI Implementation: We stay until AI is part of how the business runs

Ready to Get Your AI Model Decision Right?

400+ engagements. Clients include Zapier, Coca-Cola, Medtronic, Dataiku, and American Express.

If you are evaluating Sakana Fugu, Claude, or any other frontier model for your business, start with a conversation at Phos AI Labs.

Frequently Asked Questions

What is Sakana Fugu?

Sakana Fugu is a multi-agent AI orchestration system from Tokyo-based Sakana AI. It coordinates multiple frontier LLMs behind a single OpenAI-compatible API and behaves like one model from the outside. It launched on June 22, 2026.

What is Sakana Fugu Ultra?

Fugu Ultra is the performance-optimized variant of Sakana Fugu. It uses a deeper, fixed agent pool and is designed for complex, multi-step tasks where answer quality matters more than response speed. The current model ID is fugu-ultra-20260615.

Is Sakana Fugu open source?

No. The commercial product is a closed, hosted API. The underlying research papers (TRINITY and Conductor) are published at ICLR 2026 and available on arXiv (2606.21228), but the model itself is not downloadable or self-hostable.

How much does Sakana Fugu cost?

Subscription plans start at $20/month and go up to $200/month. All tiers include both Fugu and Fugu Ultra. Pay-as-you-go pricing for Fugu Ultra is $5 input / $30 output per 1M tokens, rising to $10 / $45 above 272K context.

What is the Sakana Fugu API base URL?

The API base URL is https://api.sakana.ai/v1. Fugu is fully OpenAI-compatible, so any client that accepts a custom base URL will work without code changes.

What models does Sakana Fugu use internally?

The pool includes Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro, among others. Anthropic’s Fable 5 and Mythos Preview are excluded from the pool due to US export control restrictions.

Can I use Sakana Fugu with Cursor or Codex?

Yes. Set the base URL to https://api.sakana.ai/v1, use your Sakana API key, and select fugu or fugu-ultra-20260615 as the model. No other changes are needed in either tool.

Is Fugu better than Fable 5 or Mythos?

On most benchmarks in Sakana’s June 2026 evaluation, Fugu Ultra leads Opus 4.8 and GPT-5.5. However, Fable 5 scores 86.0 on SWE-Bench Pro versus Fugu Ultra’s 73.7. Fable 5 is not in Fugu’s agent pool, so the two are not in a direct competitive relationship on those benchmarks.

Does Sakana Fugu have a free tier?

There is no permanent free tier. Subscribers who sign up before July 31, 2026, receive a free second month at their initial plan tier.

What is Fugu-Cyber?

On July 21, 2026, Sakana launched Fugu-Cyber, a separate gated endpoint focused on cyber defense. It reports 86.9% on CyberGym and 72.1% on CTI-REALM. It requires verification to access and is separate from the standard Fugu and Fugu Ultra products.

Can I run Sakana Fugu locally?

No. Fugu does not have a local or self-hosted version. It runs exclusively as a hosted API. There is no Ollama integration or downloadable model weight available at launch.

All benchmark figures are Sakana AI vendor-reported as of June 2026, with competing model baselines provider-reported. Independent third-party evaluation is pending. Pricing reflects Sakana’s published rates as of July 2026 and is subject to change.