> Category: LLM Gateway / Multi-model entry (orchestration) · Date: 2026-06-22 · Source: Sakana AI official blog, technical report arXiv 2606.21228
The Announcement
On June 22, 2026, Tokyo-based AI company Sakana AI officially launched Sakana Fugu and Sakana Fugu Ultra — a flagship product line that packages an entire multi-agent orchestration system as a single foundation model. The startup, backed by NVIDIA, Google, and other majors, is using "multi-agent orchestration" as its differentiated route to position against Anthropic's flagship models.
Core product positioning:
- Users access both models through a single API endpoint
- OpenAI-compatible interface — any app built on 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, synthesis — all the complexity of multi-agent systems is encapsulated inside the model, invisible to caller code
- On the most demanding engineering, science, and reasoning benchmarks, shoulder-to-shoulder with Anthropic Fable 5 and Mythos Preview
- Deliberately does not call Fable 5 / Mythos Preview (not publicly accessible) — itself a key selling point
- Consistently outperforms Gemini 3.1 Pro (high), Opus 4.8 (max), and GPT 5.5 (xhigh) across 6 real-world scenarios: AutoResearch, Rubik's Cube, Mechanical Design, Japanese Handwriting Analysis, One-Shot Chess, and Financial Time Series Prediction
- Key "AI sovereignty" narrative: eliminating single-point dependency through multi-vendor orchestration — when a vendor restricts access, Fugu routes around it automatically
- Fugu is itself a language model, specifically trained to "coordinate" — deciding when to delegate, how agents communicate, and how to synthesize results into a single reliable answer
- Key techniques come from two ICLR 2026 papers:
- Trinity: An Evolved LLM Coordinator (arXiv 2512.04695)
- Conductor: Learning to Orchestrate Agents in Natural Language (arXiv 2512.04388)
- Self-recursive invocation: Fugu can call any LLM in the agent pool, including instances of itself, recursively
- The agent pool is fully replaceable — users can exclude specific agents for data privacy or compliance reasons
- Fugu Ultra's agent pool contains no Fable 5 or Mythos Preview — restricted by export controls, which Sakana leverages to make "AI sovereignty" a core selling point
- Subscription tiers + pay-as-you-go
- ~500 beta testers reported extremely strong persona stability over long sessions — other models "drift," Fugu doesn't
- Entry: sakana.ai/fugu/ + console.sakana.ai
- For AI app developers: one API replacing multi-vendor integrations and hand-written orchestration; OpenAI compatibility plus multi-agent encapsulation means a minimal developer experience.
- For sovereignty/export-control-sensitive customers (finance, government, healthcare, critical infrastructure): a productized answer avoiding single-vendor lock-in and US export-control exposure — likely the largest adopter group 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 Japanese AI: proof that the ecosystem delivers products, business models, and geopolitical positioning — not just papers.
- For the LLM gateway space: gateways must answer how much moat remains against model-layer orchestration.
- For Anthropic/OpenAI: their customers may actively seek non-dependent alternatives, forcing a response on the "AI sovereignty" narrative.
- Rigor of the "shoulder-to-shoulder" claim: baselines are vendor self-reported scores, which tend to run high. Independent benchmark verification over 3–6 months is needed.
- Excluding Fable 5 / Mythos Preview cuts both ways: it's a sovereignty selling point but a hard performance ceiling — on tasks where those models remain strongest, Fugu Ultra will lose.
- Marginal gains of multi-agent coordination: single LLMs often win on simple tasks; Fugu's accuracy in judging "when to orchestrate" determines real-world consistency.
- Depth of OpenAI compatibility: whether it's 100% or ~80% compatible (tool calling, structured output, etc.) determines drop-in replacement viability.
- Engineering stability of recursive self-calls: context management, token cost, loop detection, and error recovery must all be solved internally.
- Commercial execution speed: whether a 2023-vintage startup can sustain a subscription + usage-based model is an open question for Japanese AI's product pipeline.
Fugu vs Fugu Ultra:
| Dimension | Fugu | Fugu Ultra | |---|---|---| | Positioning | Default choice for daily work | Maximum quality for hard multi-step problems | | Performance & latency | Balanced strong performance + low latency | Maximizes answer quality | | Agent pool | Shallower pool of experts | Deeper pool of experts | | Use cases | Codex coding / code review / chatbots / interactive services | AI research / paper reproduction / cybersecurity analysis / literature & patent surveys |
Fugu Ultra highlights:
Technical foundation:
Commercialization:
Deep 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 to some customers, making single-vendor dependency a tangible risk. Fugu Ultra positions directly against this pain point: multi-vendor orchestration, explicit refusal to call the two restricted flagships, and a product-level promise to automatically bypass restricted vendors. In effect, "AI sovereignty" is packaged as a buyable, auditable API.
2. Packaging orchestration into a model is a deep design choice. Multi-agent systems have historically been built at the application layer (LangGraph, AutoGen, CrewAI — "LLMs calling LLMs" in app code). Fugu instead trains coordination itself as a language model: callers see one OpenAI-compatible API; coordination is learned (per Trinity/Conductor) rather than hand-coded; and recursive self-invocation enables arbitrarily deep task trees. This "model orchestrating models" pushes multi-agent from application-layer hack to native model-layer capability — an important 2026 Q2 paradigm for the agent industry.
3. What "shoulder-to-shoulder" means in engineering terms. Sakana claims parity with Fable 5 and Mythos Preview, with baselines drawn from vendors' self-reported scores (using mini-swe-agent scaffolding for SWE tasks), and highlights six scenarios chosen where Fugu Ultra excels. Sakana doesn't need Fugu Ultra to be the absolute best — it needs it to hold up on tasks customers care about. Beta feedback (finding 20+ issues in code reviews where other tools found 3) suggests that goal is largely met.
4. Relationship to OpenRouter, Portkey, LiteLLM. Generic LLM gateways route by cost/latency/quality; Fugu is a model-layer-orchestrated multi-agent system that adapts to task difficulty. Its real competitors are not gateways but Anthropic/OpenAI/Google flagships: one product matching multiple vendors' top models, resilient to export controls, framed as AI sovereignty.
5. Strategic comeback for Japanese AI. Sakana AI was founded in 2023 by David Ha (ex-Google Brain) and Llion Jones (co-author of the Transformer paper). Fugu Ultra represents a product-level resurgence for Japan's AI sector: internationally recognized research (two ICLR 2026 papers), capital backing from NVIDIA and Google, and geopolitical positioning around "Japanese AI sovereignty" amid tightening export controls.
Why It Matters
Risks and Open Questions
Sources: https://sakana.ai/fugu-release/ · https://arxiv.org/abs/2606.21228 · https://github.com/SakanaAI/fugu/