> Category: LLM Gateway / Multi-model entry (orchestration) · Date: 2026-06-22 · Source: Sakana AI official blog, technical report arXiv 2606.21228
What Happened
On June 22, Sakana AI, a Tokyo-based AI company backed by NVIDIA and Google, officially launched Sakana Fugu and Sakana Fugu Ultra—a flagship product line that packages an entire multi-agent orchestration system into a single foundation model.
Core product positioning:
- Users access both models through one API endpoint
- OpenAI-compatible interface—any app using the OpenAI SDK can switch seamlessly
- Simple tasks are solved directly; for complex tasks, Fugu assembles and coordinates a team of expert models internally
- Model selection, task delegation, verification, and synthesis—all the complexity of multi-agent systems is encapsulated inside the model, invisible to calling code
- Claimed to be shoulder-to-shoulder with Anthropic's Fable 5 and Mythos Preview on the most demanding engineering, science, and reasoning benchmarks
- Deliberately does not call Fable 5 / Mythos Preview (not publicly accessible)—itself a key selling point
- In six real-world scenarios (AutoResearch, Rubik's Cube, Mechanical Design, Japanese Handwriting Analysis, One-Shot Chess, Financial Time Series Prediction), it reportedly consistently outperforms Gemini 3.1 Pro (high), Opus 4.8 (max), and GPT 5.5 (xhigh)
- AI sovereignty narrative: multi-vendor orchestration eliminates single points of dependency—when a vendor restricts access, Fugu automatically routes around it
- Fugu is itself a language model, specifically trained to coordinate—deciding when to delegate, how agents communicate, and how to synthesize results into one reliable answer
- Key techniques come from two ICLR 2026 papers: Trinity (An Evolved LLM Coordinator, arXiv 2512.04695) and Conductor (Learning to Orchestrate Agents in Natural Language, arXiv 2512.04388)
- Recursive self-invocation: Fugu can call any LLM in its agent pool, including its own instances
- The agent pool is fully swappable—users can exclude specific agents for data privacy or compliance reasons
- Fugu Ultra's pool contains no Fable 5 / Mythos Preview—these models are export-control restricted, which Sakana turns into an "AI sovereignty" selling point
- Route A: application-layer orchestration—LangGraph, AutoGen, CrewAI orchestrate LLMs in application code
- Route B: model-layer orchestration—Fugu's path, training "coordination" itself as a language model
- OpenRouter / Portkey / LiteLLM = generic gateways routing by cost/latency/quality
- Fugu = a multi-agent system with model-layer orchestration, adjusting depth by task difficulty
- For app developers: one API standing in for many models—no per-task vendor APIs, no hand-written orchestration logic. OpenAI compatibility + encapsulated multi-agent = minimal developer experience.
- For sovereignty-sensitive customers (finance, government, healthcare, critical infrastructure): a productized answer to vendor lock-in and export-control exposure—likely the biggest adoption segment in H2 2026.
- For the AI agent industry: model-layer orchestration may become the paradigm for next-gen agent frameworks; expect counter-research from OpenAI, Anthropic, and Google DeepMind within 6–12 months.
- For Japan's AI industry: proof that Japan can deliver product, business, and geopolitical positioning—not just papers and researchers.
- For the LLM gateway track: OpenRouter/Portkey/LiteLLM must answer how much moat a traditional gateway retains against model-layer orchestration.
- For Anthropic/OpenAI: their customers may be actively seeking vendor-independent alternatives.
- Rigor of the "shoulder-to-shoulder" claim: baselines are vendor self-reported and often optimistic; independent reproduction needs 3–6 months of benchmarking.
- Excluding Fable 5 / Mythos Preview cuts both ways: a sovereignty selling point, but a hard performance constraint wherever those models remain strongest.
- Marginal returns of multi-agent coordination: judging precisely when to orchestrate vs. answer directly determines real-world stability.
- Depth of OpenAI compatibility: whether tool calling, function calling, and structured output are 100% or 80% compatible decides drop-in replacement viability.
- Engineering stability of recursive self-calls: context management, token cost, loop detection, and error recovery at arbitrary depth.
- Commercialization speed: whether a 2023-founded startup can sustain subscription + usage-based revenue.
Fugu vs Fugu Ultra:
| Dimension | Fugu | Fugu Ultra | |---|---|---| | Positioning | Default choice for daily work | Maximum quality for hard multi-step problems | | Performance/latency | Balanced performance + low latency | Maximized answer quality | | Agent pool | Shallower expert pool | Deeper expert pool | | Use cases | Codex coding / code review / chatbots / interactive services | AI research / paper reproduction / cybersecurity analysis / literature & patent surveys |
Fugu Ultra highlights:
Technical foundations:
Commercialization: subscription tiers + pay-as-you-go; ~500 beta users reported strong persona stability in long sessions, where other models "drift" but Fugu does not. Entry points: https://sakana.ai/fugu/ and https://console.sakana.ai
Analysis
1. "AI sovereignty" is the real story
On June 12, 2026, Anthropic announced that a US export-control directive suspended all access to Fable 5 and Mythos Preview (see https://www.anthropic.com/news/fable-mythos-access). Direct access to Anthropic's top models became unavailable for some customers, making single-vendor dependency a tangible risk. Sakana's answer: multi-vendor orchestration, an explicit commitment not to call the restricted models, and a product-level promise to automatically route around any vendor's access restrictions. This turns "AI sovereignty" from a political concept into something you can buy, use, and audit as an API.
2. Packaging orchestration into a single model is a deep design choice
Multi-agent systems have followed two routes:
Route B means the caller is fully insulated behind one OpenAI-compatible API; coordination is learned (per Trinity/Conductor) rather than hand-rule-based; and recursive self-invocation enables in principle arbitrarily deep delegation trees. "Models orchestrating models" pushes multi-agent from application-layer hack to model-native capability—an emerging paradigm of 2026 Q2.
3. What "shoulder-to-shoulder" means in engineering terms
Sakana claims parity with Fable 5 / Mythos Preview, but the baseline scores are vendor self-reported (with mini-swe-agent scaffolding for SWE tasks), and the six showcased scenarios are ones Fugu Ultra favors. Sakana doesn't need to beat Fable 5—it needs Fugu Ultra not to fall apart on tasks customers actually care about. Per beta feedback (e.g., finding "20+ issues in code review where other tools found 3"), that goal appears largely met.
4. Not an LLM gateway, but an "AI-sovereignty LLM gateway"
Fugu's real competitors are Anthropic/OpenAI/Google flagships themselves, offering "one product rivaling multiple vendors' top models + export-control resilience + AI sovereignty."
5. Strategic comeback for Japan's AI industry
Founded in 2023 by David Ha (ex-Google Brain) and Llion Jones (Transformer co-author), Sakana has ICLR 2026 papers, big-tech capital backing, and a geopolitical position: as US–China export controls intensify, "Japanese AI sovereignty" is an emerging concept. Fugu Ultra is Japan's answer: "We don't compete on a single model—we compete on multi-agent coordination plus AI sovereignty."
Why It Matters
Risks and Things to Watch
Sources: https://sakana.ai/fugu-release/ · https://arxiv.org/abs/2606.21228 · https://github.com/SakanaAI/fugu/