Fugu Ultra Explained

Fugu Ultra is a multi-agent AI system built by Sakana AI that dynamically orchestrates frontier models like Claude, GPT, and Gemini to tackle complex, multi-step problems without relying on a single vendor. Released June 22, 2026, it represents a new approach to frontier AI: instead of picking one model, Fugu Ultra automatically assigns different parts of your task to different specialized models, coordinating their answers into a unified response.

The system solves a real problem faced by teams doing demanding analytical work. Single-model AI has gaps. Fugu Ultra closes them by learning which combination of models works best for each problem type—coding, reasoning, scientific analysis, or agentic tasks. It's designed for users who need frontier-grade answers on hard problems and are willing to pay for quality over speed.

Fugu Ultra costs $5 per million input tokens and $30 per million output tokens, with subscription tiers available. It achieves competitive performance on complex reasoning and coding benchmarks while avoiding vendor lock-in and addressing geopolitical supply-chain risk in AI infrastructure.

This guide walks you through what Fugu Ultra is, how it performs against competitors, where it excels, its real limitations, and whether it fits your workflow.

What Is Fugu Ultra?

Fugu Ultra is a multi-agent AI orchestration system that sits between you and frontier models, routing requests dynamically. Sakana AI, a Tokyo-based research lab, released it on June 22, 2026. Instead of owning its own model weights, Fugu Ultra orchestrates Claude, GPT, Gemini, and other frontier models through a single unified API.

The system uses learned coordination strategies (based on two ICLR 2026 papers: TRINITY and Conductor) to decide which model should handle which part of your request. For a complex research task, Fugu might route code generation to one model, reasoning to another, and synthesis to a third—all behind the scenes. You see one answer.

Fugu Ultra is the higher-performance variant of the Sakana Fugu family. A second variant, standard Fugu, prioritizes lower latency and cost. Both are available via OpenAI-compatible API. The weights are not open-source; you access Fugu Ultra only through Sakana's hosted API.

Key Features & Capabilities

Fugu Ultra coordinates a deep pool of expert models on each request, tuned for maximum accuracy on hard problems. It has a 1-million-token context window, supports vision/image input, and enables tool use and function calling.

The system's main strength is multi-step reasoning tasks. It learns to assemble different models for different parts of your task instead of using hand-designed workflows. This makes it particularly effective for Kaggle competitions, paper reproduction, patent research, and sustained analytical work.

Fugu Ultra avoids single-vendor dependency and geopolitical supply-chain risk. Sakana markets it as achieving frontier performance without export-control exposure. Early users can also exclude specific agent models to meet compliance or procurement policies, though only the standard Fugu variant exposes agent selection; Fugu Ultra's routing is opaque.

Benchmarks & Performance vs. Competitors

Fugu Ultra's benchmarks are vendor-reported by Sakana AI. Independent third-party verification does not yet exist. On coding tasks, it scores 73.7 on SWE-Bench Pro, 93.2 on LiveCodeBench, 90.8 on LiveCodeBench Pro, and 82.1 on TerminalBench.

On reasoning and knowledge tasks, Fugu Ultra scores 95.5 on GPQA-D, 86.6 on CharXiv Reasoning, and 50.0 on Humanity's Last Exam. Sakana claims it performs comparably to leading closed models on most benchmarks.

Fable 5, released by Anthropic on June 9, 2026, achieved generally available status via Claude API and major cloud partners (AWS, Google Cloud, Microsoft). Export controls were initially applied, then lifted on June 30, 2026, restoring access. Fable 5 is not in Fugu Ultra's agent pool, so the comparison is indirect. Fugu Ultra is bound by the frontier models it orchestrates—it cannot exceed their capabilities.

Pricing & Cost

Fugu Ultra pricing is straightforward: $5 per million input tokens and $30 per million output tokens on a pay-as-you-go basis. Subscription tiers start at $20/month (standard tier), $100/month (Pro tier with 10× usage), and $200/month (Max tier with 30× usage).

At the June 2026 launch, Sakana offered a limited promotion: subscribe before July 31, 2026, and receive a free second month. This offer remains active with 30 days remaining. Input tokens are typically cheaper than output tokens because input is the request you send, while output is the AI-generated response.

The orchestration overhead is real: Fugu Ultra uses hidden routing tokens that count toward your usage. On simple tasks, you may burn more tokens than a direct API call to a single model would use. For complex multi-step problems, the routing overhead is dwarfed by quality gains. For quick, single-turn queries, orchestration cost exceeds benefit.

Where Fugu Ultra Excels

Fugu Ultra shines on sustained, multi-step analytical tasks. Real users have deployed it for ML research (running 123 automated experiments and finding best-reported results), security assessments (end-to-end penetration testing with scope awareness and no destructive actions), and code review (surfacing 20+ issues versus 3–5 from other tools).

Patent research, literature investigation, and Kaggle competitions are signature use cases. Paper reproduction is another: Fugu Ultra's ability to coordinate different expert models for literature review, experimental design, and implementation makes it well-suited to AI researchers working to verify published claims.

Fugu Ultra is best for work where correctness matters more than speed. If you need an answer to a hard problem and are willing to wait a few seconds longer, Fugu Ultra's orchestration overhead is worth it. Conversational chatbots, real-time chat applications, and quick single-turn queries are poor fits.

Limitations & Real Drawbacks

Fugu Ultra's biggest weakness is recency: it launched June 22, 2026, and independent data is sparse. All published benchmarks are Sakana-reported. No neutral third party has reproduced Sakana's claims on standardized leaderboards. Treat claims as promising but provisional.

The routing decisions are proprietary and opaque. You cannot audit which underlying model handled which part of your request. For compliance-sensitive work (medical, financial, legal), this is a serious limitation. You cannot export the reasoning chain for audit. Only the standard Fugu model allows you to exclude specific agents; Fugu Ultra does not.

Latency is higher than single-model APIs. Orchestration has overhead. On simple tasks, you pay extra tokens for nothing. Capability is jagged: early testing of a ThreeJS task reportedly returned notably worse results than GPT-5.5 alone. Fugu Ultra is bounded by the underlying models in its pool. If all frontier models struggle with a task, so will Fugu Ultra.

Availability is limited. Fugu is not widely available in the European Union due to privacy compliance issues, and Sakana has not published a timeline for EU launch. US and Asia availability is broad.

How Fugu Ultra Compares to Alternatives

Fugu Ultra vs. Claude Fable 5 (Anthropic): Fable 5 is now generally available after export controls were lifted on June 30, 2026, restoring access initially suspended on June 12. Fable 5 launched on June 9, 2026, via Claude API and major cloud partners. Sakana claims performance parity on benchmarks, but Fable 5 is not in Fugu's orchestration pool. If your workflow depends on Claude's documented support and audit trails, Fable 5 offers clearer transparency. Fugu Ultra offers vendor-neutral orchestration at the cost of auditability.

Fugu Ultra vs. GPT 5.5 (OpenAI): Fugu Ultra reports competitive coding and reasoning benchmarks against GPT 5.5 per Sakana's testing. GPT 5.5 is a monolithic model with clearer reasoning chains. If you need compliance auditing or low-latency responses, GPT 5.5 is simpler. For hard multi-step reasoning, Fugu Ultra has a structural advantage.

Fugu Ultra vs. Gemini 3.1 Pro (Google): Fugu Ultra reports higher coding and reasoning benchmark scores. Gemini 3.1 has broad capability across different task types. Fugu Ultra is optimized for orchestration and multi-step analytical tasks; Gemini is optimized for generalist capability.

The core trade-off: Fugu Ultra sacrifices simplicity and auditability for multi-model orchestration and (potentially) higher accuracy on hard problems. If you need transparency, vendor lock-in is acceptable, or you work with simple tasks, a single frontier model is better. If you do sustained analytical work and can tolerate orchestration overhead, Fugu Ultra is worth testing.

How to Access Fugu Ultra

Fugu Ultra is available exclusively via Sakana AI's hosted API. Visit sakana.ai to sign up. You can start with pay-as-you-go pricing ($5 input / $30 output per million tokens) or choose a monthly subscription.

Access is via OpenAI-compatible API endpoints, so most SDKs and tools that support the OpenAI API format work with Fugu Ultra with minimal changes. Sakana provides API documentation and code examples on their developer portal.

Currently, availability is broad in the US and Asia but limited in the European Union. If you are based in the EU, contact Sakana sales to check timeline. No open-weights release is planned; you must use the API.