Tokyo AI lab Sakana AI officially unveiled “Sakana Fugu” on June 22, an AI orchestration system centered on multi-agent collaboration, claiming it can dynamically schedule collective intelligence of multiple models to achieve performance on par with or even surpassing Anthropic’s strongest models Fable 5 and Mythos Preview on benchmark tests for coding, scientific reasoning, and more, while being completely free from export control restrictions.
Japanese AI startup unveils Sakana Fugu model
Not bigger models, but better managers.
Fugu’s core concept differs from the traditional AI development trajectory. Over the past few years, AI progress has primarily come from brute-force scaling: building increasingly large monolithic models and feeding them ever more training data. But Sakana AI has held firm to one core belief since its founding: the most powerful AI systems won’t be isolated monoliths, but collaborative ecosystems.
Fugu is itself a language model, but it’s trained to “invoke” other LLMs from a pool of agents, including calling copies of itself for recursive computation. When faced with a task, Fugu autonomously decides which model to use, whether to break the task into planning and execution phases, whether to have another agent verify the answer, and how to combine results into the final output. For developers, all of this is presented through a single OpenAI-compatible API—you call one model, but an entire team is working behind the scenes.
This technology is built on Sakana AI’s two ICLR 2026 papers thesisOn top of this: TRINITY (exploring evolutionary collaboration between thinkers, workers, and validators) and Conductor (using reinforcement learning to discover communication and collaboration strategies between agents). The key is that the system doesn’t just chain several agents in a pipeline, but learns how to construct the pipeline itself. Recursive depth becomes a parameter that can be dynamically adjusted at inference time, meaning users can freely trade off between quality and cost without any need to retrain the model.
Performance Benchmark Comparison
Sakana AI’s benchmark results show that Fugu Ultra performs at the same level as Anthropic’s Fable 5 and Mythos Preview on the industry’s most rigorous engineering, science, and reasoning benchmarks. Here are the key results:
- SWE-Bench Pro(Software Engineering): Fugu Ultra 73.7, surpassing Opus 4.8 (69.2) and GPT-5.5 (58.6)
- LiveCodeBench v6Code Generation: Fugu Ultra 93.2, ahead of Gemini 3.1 Pro (88.5)
- GPQA-DScientific Reasoning: Fugu Ultra 95.5, tied for the highest with Fugu Standard
- Humanity’s Last Exam(General Knowledge): Fugu Ultra 50.0, close to Opus 4.8 (49.8)
- TerminalBench 2.1(Terminal Operation): Fugu Ultra 82.1, leads Opus 4.8 (74.6)
Two versions: Fugu and Fugu Ultra
Sakana Fugu launches two versions. The standard Fugu is optimized for low latency and everyday use, ideal for interactive scenarios like code review and chatbots, and also allows enterprises to exclude specific agents from the pool based on compliance requirements. Fugu Ultra delivers the highest quality output for complex multi-step tasks, suited for compute-intensive workloads such as AI research, paper reproduction, cybersecurity analysis, and patent and literature surveys.
Both versions are accessible through a unified OpenAI-compatible API, allowing developers to switch between them simply by changing the API endpoint without rewriting any code. A notable technical detail is that Fugu can recursively call itself, reading its own output to determine if the collaboration strategy isn’t good enough, then initiating a correction process. The recursion depth becomes an inference-time adjustable knob, completely without any retraining.
If desired, users can directly switch to Fugu Ultra as the model backend within the Codex editor. Sakana AI also mentioned in their official announcement that Fugu performed exceptionally well in AutoResearch experiments, where an AI agent autonomously conducted 123 training experiments over 14 hours on a single H100 GPU, ultimately achieving the best average validation score across all seeds with Fugu Ultra, surpassing three frontier model baselines.
Sakana Fugu Official Website
Conclusion
Sakana Fugu’s emergence represents a paradigm shift in AI development: instead of continuing to stack larger and more expensive monolithic models, it’s about a smaller model learning to intelligently orchestrate existing resources. When Anthropic’s strongest models couldn’t be freely deployed worldwide due to export controls, a Tokyo-based lab replaced “brute force” with “orchestration,” carving out its own niche in the competition. For countries and companies facing GPU export restrictions or unable to access specific APIs, this “model router” thinking offers a pathway to bypass geopolitical barriers.
Source: KOCPC Chinese