Fugu

Fugu

Quick answer: Fugu is Sakana AI's multi-agent orchestration model, released 22 June 2026. It routes tasks across a dynamic pool of frontier LLMs — assembling and coordinating expert agents internally — and exposes everything through a single OpenAI-compatible API endpoint. Fugu balances performance with low latency, making it the default pick for everyday coding, code review, and interactive chatbot services.

At a Glance

Where Fugu leads

• Frontier-class coding performance — 92.9% on LiveCodeBench, 87.8% on LiveCodeBench Pro.
• Single-endpoint simplicity: no multi-agent plumbing required in user code.
• Provider resilience — agent pool is swappable; no single-vendor dependency.
• Compliance controls: opt specific providers or models out of the agent pool.

Where it lags

• Per-task cost depends on which underlying models activate; less predictable than a flat rate.
• Not available in EU/EEA markets (GDPR compliance in progress).
• Architecture opacity: routing decisions and model selections are not exposed by the API.

Best for: everyday coding assistance, code review, responsive chatbots, and interactive services that need frontier capability without multi-agent engineering overhead.

What Fugu Is

Fugu is the standard tier of Sakana AI's Sakana Fugu product, a multi-agent system that presents as a single foundation model. When a request arrives at the Fugu endpoint, Fugu's orchestrator — itself a language model trained to coordinate agents — decides whether to handle the task directly or to assemble and direct a team of specialist models. The user never sees that coordination layer; from outside, it behaves like a single model call.

The orchestration technology is grounded in two ICLR 2026 papers from Sakana AI: TRINITY, which trains a lightweight evolved coordinator to assign Thinker, Worker, and Verifier roles across multiple LLMs; and The Conductor, which uses reinforcement learning to discover natural-language coordination strategies. Fugu inherits these learned coordination patterns and applies them dynamically to each incoming task.

Fugu's agent pool is configurable for compliance-sensitive teams. Users can opt specific providers or models out of the pool — for example, routing entirely through on-premises or EU-approved models — while retaining the single-endpoint interface. This positions Fugu as a practical hedge against API access disruptions: if a provider restricts access, Fugu reroutes around it automatically.

Specifications

Pricing

Fugu uses a blended rate model based on which agents are active, not a fixed per-token price.

Fees are never stacked — adding more agents does not multiply the bill. You can also subscribe:

Every subscription tier includes access to both Fugu and Fugu Ultra. Source: Sakana Fugu pricing.

Context Window

Fugu has a context window that supports up to 272K tokens at standard rates, with extended context available at a higher rate. At 272K tokens this covers approximately 400+ pages of text in a single request — sufficient for full codebases, long research papers, or multi-document analysis tasks.

Fugu vs Alternatives

Fugu and Fugu Ultra both exceed the GPQA Diamond scores of individually callable frontier models, because Fugu can coordinate multiple reasoning specialists simultaneously. The practical tradeoff: Fugu's variable pricing is harder to forecast for heavy workloads, and it is unavailable in the EU/EEA.

How Fugu Performs on Real Agent Tasks

Sakana AI's beta data from nearly 500 early users — covering code review, paper reproduction, cybersecurity analysis, and patent research — consistently showed Fugu outperforming single-call frontier models on long, multi-step tasks. In an AutoResearch experiment, Fugu Ultra (the stronger sibling) drove 123 experiment iterations autonomously over 14 hours, finishing with the best mean bits-per-byte across four seeds. In a security assessment test, Fugu executed a full recon-to-report pipeline from a single scoped instruction, staying inside scope without destructive actions.

For Benchgen users, the relevant signal is reliability-under-repetition: the public benchmark scores show Fugu's ceiling, but what matters for a specific domain workflow is whether it completes that workflow consistently. Because Fugu's routing is non-deterministic and provider-dependent, Benchgen runs expose variance that single-trial vendor benchmarks cannot.

Use Fugu via API

from openai import OpenAI  # Fugu exposes an OpenAI-compatible API

client = OpenAI(
    api_key="YOUR_SAKANA_API_KEY",
    base_url="https://api.sakana.ai/v1",  # replace with actual endpoint
)

response = client.chat.completions.create(
    model="fugu",
    messages=[{"role": "user", "content": "Review this Python function for bugs..."}],
)
print(response.choices[0].message.content)

Related Models & Benchmarks

• Compare: Fugu Ultra · Claude Fable 5 · Gemini 3.1 Pro
• Benchmarks: SWE-bench Pro · LiveCodeBench · GPQA Diamond · Humanity's Last Exam

Frequently Asked Questions

Specs and scores sourced from Sakana AI's Fugu announcement and the Fugu technical report. Last updated 2026-06-22.