Is Sakana Fugu really better than Anthropic's Fable 5 and Mythos preview?

What is Sakana Fugu?

It is a multi-agent AI system that presents itself as a single model. Instead of relying on a fixed workflow, it dynamically selects, coordinates, and orchestrates specialised AI models for each task, optimising performance across coding, reasoning, research, and other complex workloads.

This means that instead of being a single traditional standalone large language model (LLM), Fugu acts as a conductor or a smart project manager that evaluates your prompt, breaks it down, and coordinates a hidden pool of frontier AI models (including GPT, Claude Opus, and Gemini) to solve the task.


Who are the people behind Sakana Fugu?

The startup, founded in 2023, is led by David Ha and Llion Jones, who was a co-author of the famous foundational Google paper, ‘Attention Is All You Need’ (2017). Tokyo-based Sakana AI closed a ¥20 billion (approximately $135 million) Series B funding round at a post-money valuation of $2.65 billion last November, according to a TechCrunch report.

How does the agentic AI system work?

Per the startup’s website, the system is built on two research papers — TRINITY and Conductor — presented at the International Conference on Learning Representations (ICLR) 2026 edition, one of the world’s top machine learning conferences.

Fugu learns how to assemble expert agents, assign roles, and coordinate collaboration patterns rather than following predefined structures. Users access the system through a single OpenAI-compatible application programming interface (API), a standardised way of accessing different models without changing coding language, while Fugu handles model routing, switching, and orchestration in the background.

The platform offers two versions. Fugu is designed for everyday use, balancing performance and latency (the delay in sending and receiving a response) for coding, code review, chatbots, and research tasks. Fugu Ultra prioritises answer quality, coordinating a larger pool of agents for demanding workloads such as paper reproduction, cybersecurity analysis and patent research.

How does it fare against competitors?

Per Sakana AI, Fugu Ultra matches or exceeds leading frontier models on several coding, reasoning, scientific, and agentic benchmarks, including SWE Bench Pro, LiveCodeBench, GPQA-D, and Humanity's Last Exam. The company argues that orchestrating multiple strong models can outperform any individual model on complex tasks.

However, Sakana Fugu is not a frontier foundation model in the same sense as Claude Opus, GPT 5.5 or Gemini 3.1 Pro. Instead, it coordinates a pool of underlying AI models. While users interact with Fugu as a single model, its performance is dependent on the models in its agent pool.

Hence, while it outperforms other foundational LLMs, the startup explicitly stated in its blogpost that users cannot see which underlying models were used to process their queries, because model selection and routing are proprietary. That strongly suggests Fugu's capabilities depend, at least in part, on access to other models rather than solely on a new foundation model they trained themselves.

How is it priced?

The model is available through both subscriptions and pay-as-you-go usage. Subscription plans range from $20 to $200 per month, while Fugu Ultra API pricing starts at $5 per million input tokens and $30 per million output tokens. Users can configure agent participation in standard Fugu for privacy and compliance needs, while Fugu Ultra uses a fixed agent pool.