Japanese artificial intelligence startup Sakana AI on June 22, 2026, launched Fugu, a multi-model orchestration system designed to coordinate multiple frontier AI models through a single API — positioning itself less as a chatbot and more as a project manager for other AIs.
The name Fugu means "pufferfish" in Japanese, while Sakana translates as "fish" — a reference the company has used to describe how simple rules among many agents can produce collective intelligence, much like a school of fish.
Rather than answering queries directly, Fugu assesses incoming tasks and routes different components — web searches, code writing, result verification — to specialized AI agents within what the company calls an "agent pool." From the user's perspective, the interface remains a single chat window; behind it, multiple models may be working simultaneously.
Not a Chatbot, Not a Router
Fugu is itself a language model trained by Sakana AI, but its core function is coordination rather than direct inference. The company describes it as "a multi-agent orchestration system delivered as a single foundation model" — packaging what enterprises would otherwise have to build themselves into something callable like a standard model via an OpenAI-compatible API.
This distinguishes Fugu from a conventional model router, which simply redirects a request to a chosen model. According to Sakana AI, Fugu decides how to decompose a task, whether multiple models need to collaborate, whether outputs need to be verified, and how to synthesize a final response. The company compares a model router to a traffic signal, and Fugu to a project manager.
The underlying architecture draws on two Sakana AI research efforts: TRINITY, which assigns separate models to thinking, execution, and verification; and The Conductor, which uses reinforcement learning to teach a system to coordinate AI agents using natural language.
How It Compares to ChatGPT and Claude
Products such as ChatGPT, Claude, and Gemini typically rely on a single model to complete an entire reasoning process. Fugu's architecture distributes that work. In a software development task, for instance, one model may write code, another may audit it for vulnerabilities, and a third may check whether the output meets the original specification — with Fugu determining the division of labor and the integration of results.
| Dimension | Conventional LLM (ChatGPT / Claude / Gemini) | Sakana Fugu |
|---|---|---|
| Task completion | Single model | Multi-model collaboration |
| Access method | Call a single model | Single API; system coordinates multiple models internally |
| Model selection | Chosen by user or developer | Automatically dispatched by system based on task |
| Scalability | Primarily by upgrading or swapping the model | Add or replace models within the agent pool |
| Core value | Single-model capability | Division of labor, verification, and integration across models |
Benchmark Claims and Caveats
Sakana AI also released Fugu Ultra, a higher-tier variant targeting complex, multi-step tasks. According to the company's own testing, Fugu Ultra achieved the following scores on several AI benchmarks:
- SWE-Bench Pro: 73.7
- TerminalBench 2.1: 82.1
- LiveCodeBench: 93.2
- LiveCodeBench Pro: 90.8
- GPQA-Diamond: 95.5
Sakana AI claims Fugu Ultra performs "shoulder-to-shoulder" with Anthropic's Fable 5 and Mythos Preview. Notably, the company stopped short of asserting that Fugu Ultra outperforms those models across the board — only that it is broadly comparable. The standard Fugu scored 92.9 on LiveCodeBench, meaning Fugu Ultra's margin over its own baseline on that test is narrow. These results are based on internal testing and have not been independently verified by third parties; in some direct comparisons, Anthropic's Fable 5 still leads.
Pricing
Sakana AI's published pricing for Fugu Ultra is $5 per million input tokens, $30 per million output tokens, and $0.50 per million cached input tokens. For contexts exceeding 272,000 tokens, pricing rises to $10 per million inputs, $45 per million outputs, and $1 per million cached inputs.
The company positions Fugu Ultra around high-complexity tasks, multi-model collaboration, and supply-chain resilience rather than low per-token cost. For simpler tasks — bulk classification, short-form rewriting, format conversion — a single-model solution is likely more cost-efficient. Fugu Ultra's value proposition is more relevant for multi-step reasoning, code analysis, cross-referencing data, or research-oriented workloads.
The Export Control Context
The timing of Fugu's launch was notable. In June 2026, the U.S. government directed Anthropic on national security grounds to suspend foreign access to frontier models including Fable 5 and Mythos 5. Anthropic complied by disabling global access to both models, triggering wider industry concern about AI supply-chain risk.
Sakana AI explicitly cited this episode in its product announcement, describing Fugu's multi-model architecture as offering resilience against such disruptions and characterizing the product as available "without the risk of export controls." The company argues that because Fugu can substitute one model for another within its agent pool, enterprises are less exposed to the sudden unavailability of any single provider.
That framing comes with limits, however. Fugu does not avoid U.S.-made models entirely, nor does it place users outside the reach of export-control regimes. The more precise claim is that its swappable architecture preserves flexibility — reducing, but not eliminating, dependency on any one vendor.
Japan's Broader AI Anxiety
Sakana AI's approach resonates with a wider anxiety in Japan about falling behind in artificial intelligence. Hisashi Matsumoto (松本剛明), Japan's digital minister, warned on June 5, 2026, that Japan risked becoming an "AI colony" if it failed to keep pace with AI development — a remark made in defense of proposed legislation that would loosen consent requirements for using sensitive data, including medical and criminal records, in AI training. The bill has drawn domestic criticism over privacy implications.
Nonetheless, the underlying concern — avoiding dependency on foreign technology in a strategically critical sector — maps directly onto what Sakana AI is selling. Rather than competing head-on with OpenAI, Anthropic, or Google on model scale, compute, or funding, Sakana AI is betting on a different axis: building the system that integrates and coordinates models, rather than the largest model itself.
If foundation models continue to commoditize, the argument goes, orchestration platforms that sit above them could become the next competitive battleground.
Key Facts at a Glance
- What it is: A multi-model orchestration system launched by Sakana AI on June 22, 2026.
- How it works: Fugu coordinates multiple large language models and AI agents, accessible via a single OpenAI-compatible API.
- Two tiers: Fugu balances performance and latency; Fugu Ultra targets high-difficulty, multi-step tasks.
- Claimed scores (Fugu Ultra): SWE-Bench Pro 73.7, TerminalBench 2.1 82.1, LiveCodeBench 93.2, LiveCodeBench Pro 90.8, GPQA-Diamond 95.5.
- Key differentiator: Reduced dependence on any single model provider through orchestrated multi-model collaboration.
- Pricing (Fugu Ultra): $5 input / $30 output per million tokens; higher rates apply beyond 272K token contexts.
- Caveat: Benchmark results are based on Sakana AI's internal testing and have not been independently verified. Enterprises should validate actual cost, performance, and reliability against their own use cases before deployment.
For years, AI competition has centered on building bigger models. Fugu represents a different premise: that the more consequential capability may lie in teaching AI to direct other AI. Whether that orchestration layer becomes the decisive frontier remains to be seen — but the launch marks a concrete step toward making it a commercial product.
Sources: Sakana AI product page, Sakana AI pricing page, Sakana Fugu technical report, Reuters (June 5, 2026), Anthropic official statement.
