What Is Sakana Fugu Ultra? Sakana AI's Top-Tier Model

Sakana Fugu Ultra is the higher-tier variant of Sakana AI's Fugu multi-agent orchestration system, tuned for maximum answer quality on hard, multi-step engineering, scientific, and reasoning problems. Announced alongside standard Fugu on June 22, 2026, it draws on a deeper expert-agent pool than the base model while staying exposed through the exact same single, OpenAI-compatible API, so callers get frontier-tier depth without managing any of the orchestration themselves.

What Sakana Fugu Ultra is

Sakana Fugu Ultra is the maximum-quality tier of Sakana AI's Fugu, a foundation model whose core ability is coordinating other AI models rather than only generating tokens itself. Where standard Fugu balances speed and quality for everyday work, Fugu Ultra is built for the hardest requests where getting the best possible answer matters more than latency or cost.

From the developer's point of view, sakana ai fugu ultra is identical to calling any single model: one prompt in, one answer out. The "Ultra" label signals a deeper pool of expert agents the system can recruit and coordinate behind the scenes, not a different API surface or integration path.

Sakana AI, a Tokyo-based lab, frames the whole Fugu line around a founding idea: "The most powerful AI systems will not be isolated monoliths, but collaborative ecosystems." Fugu Ultra is the high end of that vision, packaged so the orchestration stays invisible to the caller. You can read the company's own framing on the Sakana Fugu release announcement.

How Fugu Ultra works

Fugu Ultra works by having a language model select, delegate to, verify, and synthesize the output of other LLMs in an expert-agent pool, including recursive self-calls. The Ultra tier widens and deepens that pool so harder subproblems can be routed to more or stronger specialists before the results are merged into one reply.

In practice, Fugu Ultra reads your request, decides which experts to involve, hands them subtasks, checks the returned work, and assembles a single coherent answer. Because it supports recursive self-calls, the system can invoke itself as one of the experts when a subproblem benefits from the same orchestration logic — useful for layered, multi-step reasoning.

All of this collapses behind one OpenAI-compatible endpoint, so your application never needs to know how many models were consulted or how the work was split. The technical underpinnings are detailed in the Sakana Fugu Technical Report (arXiv:2606.21228), which describes the coordination approach shared across both tiers.

Fugu vs Fugu Ultra

The difference between Fugu and Fugu Ultra is depth versus speed: standard Fugu is tuned for balanced performance and low latency, while fugu ultra draws on a deeper expert-agent pool optimized for maximum answer quality. Both share the same orchestration design and the same OpenAI-compatible API.

Sakana AI positions standard Fugu for everyday tasks, interactive services, and code review, where responsiveness matters. Fugu Ultra is aimed at hard, multi-step engineering, scientific, and reasoning problems where the best possible answer outweighs the extra latency and cost.

Performance claims

Sakana AI reports that Fugu Ultra stands shoulder-to-shoulder with frontier models across rigorous engineering, scientific, and reasoning work. These are the company's own characterizations — vendor claims, not independently verified as of June 2026.

According to Sakana AI, the comparison set referenced for sakana fugu ultra includes Fable 5, Mythos Preview, Gemini 3.1 Pro, Opus 4.8, and GPT-5.5. The company describes Fugu Ultra as competitive with that frontier tier rather than positioning it as a clear, measured winner in its public materials.

Some press coverage goes further, framing Fugu Ultra as edging out individual frontier models on specific benchmarks. Treat those as reported claims attributed to outlets, not as verified results; as of late June 2026, broad independent testing had not caught up to the launch. You can review Sakana's own product page on the Sakana Fugu product overview.

The practical takeaway: Fugu Ultra is plausibly a frontier-class option on hard tasks, but the strongest performance statements still trace back to the vendor or early press. Run it on your own representative problems before treating any launch-window number as fact.

Access and pricing

Sakana Fugu Ultra is accessed through the same OpenAI-compatible API as base Fugu, which means tooling that already speaks the OpenAI format can target Sakana with minimal changes. Sakana AI provides this via its developer console.

On pricing, exact figures for the Ultra tier should be read as reported claims, not confirmed numbers — vendor claims and press coverage, not independently verified as of June 2026. Sakana's materials describe subscription tiers plus a pay-as-you-go plan without committing to fixed public figures in their text.

Because Fugu Ultra can recruit a deeper pool of experts per request, its effective cost and latency are typically higher and more variable than standard Fugu. 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 Ultra 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.

The deeper expert pool that makes Fugu Ultra strong also makes it less predictable. When it recruits and coordinates many models for a hard request, latency and cost can swing more than with a single fixed model — a real problem for interactive products with strict response budgets.

Vendor benchmark claims, including any reported wins over frontier models like GPT-5.5 and Opus 4.8, remain unverified by neutral third parties as of June 2026. Until reproducible independent results appear, treat the "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, documented track record, deterministic latency, or audited benchmark numbers, Fugu Ultra may not be the right bet yet. For research, hard one-off problems, and tasks where you can A/B test outputs, it is far easier to justify trying it — and dropping back to standard Fugu when speed matters more than depth.

Related Guides

• Sakana Fugu vs Claude Fable 5
• What Is Sakana Fugu?
• What Is Claude Fable 5?
• Fable 5 and Mythos 5

FAQ

What is Sakana Fugu Ultra?

Sakana Fugu Ultra is the higher-tier variant of Sakana AI's Fugu multi-agent orchestration system, announced June 22, 2026. It uses a deeper expert-agent pool than standard Fugu, tuned for maximum answer quality on hard, multi-step problems, and is exposed through the same single OpenAI-compatible API.

What is the difference between Fugu and Fugu Ultra?

Standard Fugu is tuned for balanced performance and low latency on everyday tasks, interactive services, and code review. Fugu Ultra draws on a deeper expert-agent pool optimized for maximum answer quality on hard engineering, scientific, and reasoning problems, trading some speed and cost for depth.

Is Sakana Fugu Ultra better than Claude Fable 5?

Sakana AI reports that Fugu Ultra stands shoulder-to-shoulder with frontier models including Fable 5, Mythos Preview, Gemini 3.1 Pro, Opus 4.8, and GPT-5.5. These are vendor claims, not independently verified as of June 2026, so it is best to test both on your own workloads rather than rely on launch-window numbers.

How do you access Sakana AI Fugu Ultra?

Sakana AI Fugu Ultra is accessed through the same OpenAI-compatible API as standard Fugu via Sakana's developer console, so existing OpenAI-format tooling can point at it with minimal changes. Exact pricing should be treated as reported claims and verified directly in the Sakana console.

Who makes Sakana Fugu Ultra?

Sakana Fugu Ultra is made by Sakana AI, a Tokyo-based AI lab. The model's coordination approach is described in the Sakana Fugu Technical Report (arXiv:2606.21228), which covers both the standard and Ultra tiers.