Sakana AI’s Fugu: Multi-Agent Orchestration as a Single Model

Sakana AI has released Fugu, a sophisticated multi-agent orchestration system packaged as a single language model. Accessible through one OpenAI-compatible API, Fugu either handles tasks directly or intelligently breaks them down into subtasks, delegates them across a pool of specialized models (including frontier systems from Anthropic, OpenAI, and strong open-weight options like Qwen, GLM, and DeepSeek), and synthesizes the results.

This “orchestrator-as-a-model” approach delivers frontier-level performance without forcing developers to manage complex multi-agent frameworks themselves. It also provides built-in resilience against single-vendor dependency and export controls.

How Fugu Works

Fugu is itself a trained language model specialized in learned model orchestration.

Sakana AI’s Fugu: Multi-Agent Orchestration as a Single Model Drawing from Sakana’s research papers (including Trinity and Conductor), it decides when to solve a problem on its own versus when to assemble and coordinate a team of expert agents. It handles model selection, delegation, verification, communication between agents, and final synthesis internally.

From the outside, it feels like calling any other LLM. On the inside, it routes tasks dynamically across a swappable pool of models. You can opt out of specific agents in the standard Fugu version for privacy or compliance reasons. The system improves over time as newer models are added to the pool.

This design offers practical advantages:

Automatic tool selection — It picks the best model or combination for each subtask.
Bypassing restrictions — It can route around export controls or access limitations (as seen with models like Fable/Mythos).
Resilience — No single point of failure or vendor lock-in.

Two Versions for Different Needs

Sakana offers two variants through the same API:

Fugu — Balanced for speed and everyday use. Ideal for coding assistance, code review, chatbots, and interactive tools. Lower latency makes it suitable for real-time applications.
Fugu Ultra — Tuned for maximum quality on complex, multi-step tasks. It coordinates a deeper pool of agents and excels in areas like AI research, cybersecurity analysis, scientific workflows, patent investigations, and long-horizon reasoning.

Fugu Ultra stands shoulder-to-shoulder with leading frontier models on rigorous benchmarks (e.g., strong results on SWE-Bench Pro, TerminalBench, GPQA Diamond, and more), while avoiding reliance on any single restricted model.

Real-World Performance and Examples

Early beta users and Sakana’s own experiments highlight Fugu’s strengths in agentic, long-running workflows:

Autonomous ML research — Improved training recipes through iterative experimentation (outperforming several frontier models in controlled tests).
Financial time series prediction — Achieved higher returns in simulated trading scenarios.
Mechanical design (CAD) — Generated functional models (e.g., a working mechanical iris) where single models often failed on physical logic.
Code generation and review — Produced more comprehensive bug detection and efficient solvers (e.g., Rubik’s Cube solver in pure Python).
Cybersecurity assessments — Conducted end-to-end analysis including recon, vulnerability checks, and reporting.
Other tasks — Blindfold chess, classical Japanese text analysis, and more.

In community tests, such as building a full live trading terminal with frontend, backend, and real-time market data, Fugu Ultra delivered polished, feature-rich results comparable to top models.

The Cost Trade-Off

Performance comes at a price in token consumption. Fugu’s internal orchestration often generates hidden cascades of model calls, which can significantly inflate usage.

In one detailed community benchmark building a complete trading desk application:

Fugu Ultra: ~22,225 tokens → $0.51;
GLM-5.2: ~13,677 tokens → $0.03;
Other frontier models fell in between.

This represents roughly a 17x cost difference compared to GLM-5.2 for similar (or slightly superior in polish) output quality.

Sakana’s official pricing reflects the orchestration overhead:

Subscriptions — $20/month (Standard), $100/month (Pro), $200/month (Max) — all include access to both Fugu and Fugu Ultra with usage caps.
Pay-as-you-go — Fugu Ultra is priced at $5 per million input tokens and $30 per million output tokens (higher for very long contexts). Billing is based on the highest-tier model used in the orchestration (no stacking of fees).

The company emphasizes monitoring real-time token usage, but users should expect higher costs on complex, multi-step tasks compared to single-model inference.

Access and Integration

Fugu is available via API key generated in the Sakana console (console.sakana.ai). There is no native web chat interface on the product site.

You plug the key into compatible tools such as:

OpenRouter;
Codex;
Any OpenAI-compatible client or framework.

It is not currently available in the EU/EEA due to regulatory compliance.

Why Fugu Matters

Fugu represents a meaningful shift toward collective intelligence over monolithic models.

By treating orchestration as a learned capability rather than manual engineering, Sakana delivers:

Frontier performance with greater flexibility and sovereignty.
A hedge against geopolitical or vendor-specific risks.
Simplified developer experience (one API, automatic coordination).

The main caveats are cost predictability on heavy workloads and the opaque nature of which underlying models are being routed to in any given call.

For developers and teams working on complex agentic projects, research automation, or secure/multi-vendor workflows, Fugu offers a compelling new option.

For simpler or highly cost-sensitive tasks, lighter single models (or cheaper open-weight alternatives) may still be more practical.

As AI systems become more agentic and multi-model ecosystems mature, tools like Fugu point toward a future where the “model” is increasingly a smart coordinator rather than a single brain.

Official links:

Announcement: https://sakana.ai/fugu-release/;
Product page: https://sakana.ai/fugu/;
Console for API access: https://console.sakana.ai.

The space is moving fast — Fugu is one of the clearest commercial expressions yet of learned multi-agent orchestration.