What Is Sakana Fugu? Japan's Multi-Agent AI Model

Sakana Fugu is a multi-agent orchestration system from Sakana AI, a Tokyo-based AI lab, that is packaged and exposed as a single foundation model: you send one request to one API endpoint, and Fugu decides on its own whether to answer directly or assemble and coordinate a team of expert models behind the scenes. It was announced on June 22, 2026, and is offered in two variants, Fugu and Fugu Ultra, with Sakana AI reporting that Fugu Ultra stands shoulder-to-shoulder with leading frontier models.

What Sakana Fugu is

Sakana Fugu is a foundation model whose core ability is coordinating other AI models rather than only generating tokens itself. Announced by Sakana AI on June 22, 2026, it is designed so the developer experience is identical to calling any single model: one prompt in, one answer out.

The difference is what happens inside. For a simple request, Fugu may answer directly. For a harder one, it can recruit a team of specialized models, hand them subtasks, check their work, and merge the results into a final response, all without the caller managing any of that machinery.

Sakana AI frames this around a founding principle: "The most powerful AI systems will not be isolated monoliths, but collaborative ecosystems." Fugu is the company's attempt to make that idea usable through a single, conventional API. You can read the company's own framing on the Sakana Fugu release announcement.

How the orchestration works

The mechanism behind Sakana Fugu is a language model that has been trained to call other LLMs in an agent pool and manage the entire delegation loop itself. Sakana AI describes the model as handling model selection, delegation, verification, and synthesis internally.

In practice, that means Fugu reads your request, decides which experts (if any) to involve, routes subtasks to them, evaluates the returned answers, and assembles a single coherent reply. Crucially, the system supports recursive self-calls, so Fugu can invoke itself as one of the experts when a subproblem benefits from the same orchestration logic.

Because all of this is collapsed behind one OpenAI-compatible endpoint, your application does not need to know how many models were consulted or how the work was split. The orchestration is internal, which keeps integration simple while the heavy lifting happens server-side. The technical underpinnings are detailed in the Sakana Fugu Technical Report (arXiv:2606.21228), which builds on two ICLR 2026 papers, "Trinity: An Evolved LLM Coordinator" and "Learning to Orchestrate Agents."

Fugu vs Fugu Ultra

Sakana Fugu ships in two versions that share the same orchestration approach but target different needs. Fugu is tuned for balanced performance and low latency, while Fugu Ultra is optimized for maximum answer quality on the hardest problems.

Sakana AI positions standard Fugu for everyday tasks, interactive services, and code review, where responsiveness matters. Fugu Ultra draws on a deeper expert agent pool and is aimed at hard, multi-step problems where getting the best possible answer outweighs speed or cost.

Performance claims

Sakana AI reports that Fugu Ultra "stands shoulder-to-shoulder with leading models like Fable 5 and Mythos Preview" across rigorous engineering, scientific, and reasoning benchmarks. These are the company's own characterizations, not independently confirmed results.

According to Sakana AI, the baseline comparisons referenced in its materials include Gemini 3.1 Pro, Opus 4.8, and GPT-5.5. The company describes Fugu Ultra as competitive with that frontier tier rather than publishing a single headline score in its public text.

Some press reports go further, claiming Fugu Ultra surpasses GPT-5.5 and Opus 4.8 on SWE-Bench Pro and TerminalBench. Treat those as reported claims attributed to the outlets, not as verified benchmark wins; as of late June 2026, broad independent testing has not caught up to the launch. You can review Sakana's own product page on the Sakana Fugu product overview.

Sakana AI also notes a beta cohort of roughly 500 early users spanning automated research, code review, cybersecurity analysis, paper reproduction, and patent investigation, which suggests the model is being trialed on demanding, real-world workloads.

Pricing and access

Sakana Fugu is accessed through an OpenAI-compatible API, which means existing tooling that already speaks the OpenAI format can point at Sakana with minimal changes. Sakana AI provides this via its developer console at console.sakana.ai.

On pricing, Sakana's official page describes subscription tiers plus a pay-as-you-go plan but did not list exact figures in its text. Press coverage from GIGAZINE cites reported tiers of roughly $20 per month for Standard, $100 per month for Pro, and $200 per month for Max.

Those numbers should be read as reported figures attributed to press, not confirmed pricing. Anyone planning a budget should verify current rates directly in the Sakana console before committing, since launch-window pricing often shifts.

Limitations and Tradeoffs

Sakana Fugu is brand new, having launched on June 22, 2026, so independent benchmarking is still thin and most performance claims trace back to the vendor or early press. That alone is a reason for caution if you are making a high-stakes platform decision in mid-2026.

Multi-agent orchestration also introduces variability. When Fugu recruits and coordinates several models for a hard request, latency and cost can be harder to predict than with a single fixed model, and that variance matters for interactive products with strict response budgets.

Vendor benchmark claims, including the reported wins over GPT-5.5 and Opus 4.8, remain unverified by neutral third parties. Until reproducible independent results appear, it is reasonable to treat Sakana AI's "shoulder-to-shoulder with Fable 5" positioning as a marketing claim worth testing on your own tasks.

When should you hold off? If you need a model with a long, well-documented track record, deterministic latency, or audited benchmark numbers, Fugu may not be the right bet yet. For experimentation, research, and tasks where you can A/B test outputs, it is far easier to justify trying it.

Related Guides

• Sakana Fugu vs Claude Fable 5
• What Is Claude Fable 5?
• Best Cheap AI Models in 2026
• Best Claude Models in 2026

FAQ

What is Sakana Fugu?

Sakana Fugu is a multi-agent orchestration system from Sakana AI, announced June 22, 2026, that is exposed as a single foundation model. You call one API endpoint, and Fugu decides whether to answer directly or assemble and coordinate a team of expert models internally.

How much does Sakana Fugu cost?

Sakana AI's official page describes subscription tiers plus a pay-as-you-go plan but did not publish exact figures in its text. Press reports (GIGAZINE) cite roughly $20 per month for Standard, $100 per month for Pro, and $200 per month for Max; verify current pricing in the Sakana console.

Is Sakana Fugu better than Claude Fable 5?

Sakana AI reports that Fugu Ultra stands shoulder-to-shoulder with leading models like Fable 5, and some press reports claim wins on specific benchmarks. As of June 2026 these claims are not independently verified, so it is best to test both on your own workloads rather than rely on launch-window numbers.

What is the difference between Fugu and Fugu Ultra?

Fugu is tuned for balanced performance and low latency on everyday tasks, interactive services, and code review. Fugu Ultra uses a deeper expert agent pool and is optimized for maximum answer quality on hard, multi-step engineering, scientific, and reasoning problems.

Who makes Sakana Fugu?

Sakana Fugu is made by Sakana AI, a Tokyo-based AI lab. The model is grounded in the company's Sakana Fugu Technical Report (arXiv:2606.21228) and two ICLR 2026 papers on LLM coordination and agent orchestration.