Japanese startup Sakana AI announced Fugu on Monday (the 22nd), an AI model that achieved results equivalent to Anthropic's Fable 5 and Mythos 5 on the industry's most demanding benchmarks. The difference lies in the architecture: instead of training a single giant model, Fugu orchestrates several smaller language models acting as collaborative agents. This multi-agent approach does away with billions of pre-trained parameters and cuts costs, while also sidestepping export control restrictions imposed by the United States.
TL;DR: Sakana Fugu uses multi-agent orchestration and runtime scalability to compete with giants such as Fable 5 and Mythos 5, offering cutting-edge performance without relying on a single vendor or on huge volumes of trained parameters.
The proposal arrives at a strategic moment. The US block on Anthropic's technologies cut off access for foreign users and reignited the debate over concentration of power in the global AI race. For product teams and technology managers, the Japanese model represents a viable alternative, and a conceptually different one, to the "bigger is better" strategy that dominates the industry.
How does Sakana Fugu's multi-agent orchestration model work?
Fugu is not a single colossal model. It's an orchestration system that coordinates several smaller language models, each specialized in specific functions.
When it receives a request, Fugu assesses its complexity. Simple tasks can be solved directly. Multi-step tasks, such as data analysis followed by code generation and report writing, are delegated to specialized agents. The orchestrator selects the most suitable model for each subtask, gathers the results and delivers the final answer.
This strategy brings two practical gains:
- Runtime scalability: increases reasoning capacity on demand, without retraining giant models.
- Modularity: if a specific model becomes unavailable (because of geopolitical restrictions, for example), you simply swap it for another compatible one. There's no dependence on a single vendor.
Sakana says this architecture makes it possible to avoid training giant models, cutting infrastructure costs and development time. For companies that want to adopt AI without billion-dollar budgets, this approach decentralizes the game.
Sakana Fugu vs. Fable 5 and Mythos 5: what changed in the performance comparison?
The startup reported that Fugu matched the performance of Anthropic's Fable 5 and Mythos 5 on the industry's most rigorous benchmarks, tests that measure complex reasoning, multi-step tasks and coordination across long contexts.
| Criterion | Fugu (Sakana) | Fable 5 / Mythos 5 (Anthropic) |
|---|---|---|
| Architecture | Multi-agent orchestration | Single model, billions of parameters |
| Vendor dependence | Low (interchangeable agents) | High (proprietary model) |
| Training cost | Reduced (no giant models to train) | High (massive infrastructure) |
| Export controls | No restrictions (Japanese technology) | Blocked outside the US |
| Scalability | On demand (at runtime) | Fixed (pre-trained parameters) |
The differentiator isn't just in the numbers; it's in the strategic architecture. While models like Fable 5 concentrate intelligence in a single system trained on trillions of tokens, Fugu distributes reasoning across specialized agents. This reduces technology lock-in and makes it easier to adapt to local contexts.
For product teams, the lesson is clear: comparable performance no longer requires a race toward endless parameters. Collective intelligence can be as effective as a monolithic model, with operational advantages.
Why does the multi-agent approach matter for companies outside the US?
The US export control block cut off global access to Anthropic's technologies. The measure, although initially aimed at foreign users, ended up affecting everyone, creating a strategic vacuum for companies that relied on Fable 5 or Mythos 5 in their operations.
Sakana AI positioned Fugu as a direct answer to this problem:
"Collective intelligence is the practical protection against this concentration of power. Because Fugu orchestrates an underlying set of interchangeable agents, it simply routes around vendor restrictions."
In practice, this means:
- Technological sovereignty: companies in markets outside the US (including Brazil, Europe and Asia) gain an alternative without the risk of geopolitical sanctions.
- Operational resilience: if a model is blocked or discontinued, the orchestrator replaces the agent without interrupting operations.
- Predictable cost: with no need to retrain billions of parameters at every update, the infrastructure bill goes down.
For technology managers evaluating custom AI vendors, this architecture reduces dependence and increases control. It's especially relevant in regulated sectors (finance, healthcare, government), where vendor lock-in is a strategic risk.
What are the limits and challenges of multi-agent orchestration?
Sakana's strategy is promising, but not free of trade-offs. Multi-agent orchestration brings its own complexity that product teams need to consider:
Coordination latency: each task delegated to different agents adds overhead. In real-time scenarios (customer service chatbots, voice assistants), this latency can affect the user experience. Monolithic models, since they already have everything "in-house", deliver a response in a single hop.
Integration cost: orchestrating multiple models requires abstraction layers, robust APIs and fallback logic. Smaller teams or companies without AI expertise may face a steeper learning curve than if they simply plugged in a single model through a ready-made API.
Coordination quality: final performance depends on how well the orchestrator distributes tasks. If the delegation logic is poorly calibrated, the system may pick the wrong model for a subtask, degrading the final answer. Single models avoid this problem by centralizing reasoning.
Benchmarks aren't production: results on traditional benchmarks are indicative, but they don't always reflect performance in real business cases. Companies should test Fugu on their own data and workflows before migrating.
Bottom line: the multi-agent approach is powerful when implemented well, but it demands technical maturity and rigorous testing. It's not a silver bullet; it's a strategic tool for those who want control and resilience.
How do you evaluate whether a multi-agent architecture makes sense for your use case?
Not every application benefits from multi-agent orchestration. The choice depends on the type of task, data volume and degree of control you want. Questions that help you decide:
- Are your tasks multi-step and specialized? If the workflow involves classifying, processing, generating and validating information in sequence, orchestration is worth it. Example: a contract analysis pipeline (entity extraction → summary → generation of revised clauses).
- Do you need to switch vendors quickly? If relying on a single model creates risk (geopolitical, financial, technical), multi-agent offers a way out. Ideal for global companies or those in regulated markets.
- Is training cost a barrier? If retraining giant models at every update is beyond your budget, coordinating smaller agents lowers the infrastructure bill.
- Is latency critical? In real-time applications (voice assistants, automated checkout), single models still have the edge. Orchestration adds milliseconds or seconds, depending on how complex the delegation is.
At Agência Rollin, we work with clients evaluating custom AI agents for internal workflows: customer service automation, technical content generation, product data analysis. The multi-agent architecture proves effective when the use case calls for flexibility, specialization and risk control. For more straightforward applications (a simple FAQ, binary classification), a single plug-and-play model delivers results with less overhead.
Key takeaways on Sakana Fugu and multi-agent orchestration
- Sakana Fugu reaches Fable 5 and Mythos 5 performance by orchestrating multiple specialized models, not a single giant system.
- Runtime scalability increases reasoning capacity on demand, without relying on billions of pre-trained parameters.
- The architecture sidesteps export controls and reduces vendor lock-in, which is critical for companies outside the US.
- Trade-offs include coordination latency and integration cost, which require technical maturity from the team.
- Multi-agent orchestration is ideal for multi-step, specialized tasks that demand resilience, but it doesn't replace single models in every scenario.
What does the Japanese model signal for the future of enterprise AI?
Fugu's launch marks a turning point: collective intelligence as a viable alternative to the race for ever-larger models. While the giants invest billions in chips, data centers and trillions of tokens, Sakana is betting on modular architecture, smart coordination and distributed reasoning.
For companies, this opens up possibilities:
- Vendor diversification: orchestrating agents from different sources reduces the risk of disruption.
- Local customization: specialized regional models (trained on local languages, laws and contexts) can be integrated into the system without retraining everything from scratch.
- Democratized innovation: smaller teams, without a big tech budget, can compete with smart architectures instead of brute computing force.
The challenge lies in execution. Orchestration requires careful design, rigorous testing and the ability to adjust the delegation logic as the business context evolves. It's not an off-the-shelf solution; it's a strategic platform that calls for ongoing technical curation.
At Agência Rollin, we help companies evaluate when multi-agent architectures make sense and how to implement them pragmatically, without relying on hype. The right choice isn't "single model vs. multi-agent" but which approach solves the specific problem with the best cost-benefit and the lowest risk.
