Sakana AI, the prominent Tokyo-based research and development laboratory, announced the general availability of Sakana Fugu. By altering the paradigm of the industry, Fugu presents an intelligent orchestration framework for multiple agents packed within one base model. This framework is capable of intelligently navigating through the collection of interchangeable large language models (LLMs) by means of one API endpoint, thus providing a useful defense against the risk of being dependent on one vendor only.
Engineering Collective Intelligence
Rather than functioning as a standard standalone foundation model, Fugu operates as a highly trained neural orchestrator. When an enterprise sends a request to the single OpenAI-compatible endpoint, Fugu’s underlying architecture dynamically evaluates the problem. For straightforward prompts, the system resolves the task independently to preserve speed and minimize computational overhead. For intricate, long-horizon challenges, it autonomously recruits, assigns roles to, and coordinates a tailored team of specialized expert models from its integrated pool.
This learned coordination mechanism marks a distinct departure from traditional, rigidly hard-coded multi-agent frameworks. Rooted heavily in Sakana AI’s acclaimed ICLR 2026 research papers TRINITY and The Conductor Fugu does not rely on hand-designed operational workflows. Instead, it discovers and executes non-obvious, deeply optimized collaboration, verification, and response-synthesis patterns across entirely swappable agent groups. This fluid routing capability ensures that if a specific model provider encounters downtime or introduces access restrictions, Fugu seamlessly routes workflows around the disruption to maintain enterprise continuity.
Also Read: AWS Expands Amazon Bedrock AgentCore Capabilities to Empower AI Agents with Broader Knowledge and Continuous Learning
Two Tiers, One Unified API
To accommodate varying enterprise workloads, Sakana AI has deployed the system in two distinct commercial tiers, accessible through a single, integration-ready endpoint:
- Sakana Fugu: Optimized to deliver a precise balance between high-level capability and low-latency execution. It serves as an ideal default architecture for interactive chatbots, responsive customer tools, and automated software development environments like Codex.
- Fugu Ultra: Tuned explicitly for maximum quality, absolute accuracy, and granular depth on complex, multi-step reasoning problems. Fugu Ultra coordinates a significantly deeper pool of specialized agents to master heavy workloads, including autonomous AI research, academic paper reproduction, literature and patent investigations, and deep cybersecurity threat analysis.
The system ensures rigorous data governance and flexible architecture for enterprise implementations. Enterprises that need to adhere to stringent policies on data privacy, compliance, or sovereignty can easily exclude certain model providers or particular agents from the operations of the Fugu system.
Industry-Leading Benchmarks
Across rigorous engineering, scientific, and reasoning evaluations, Sakana AI‘s flagship orchestration tier demonstrates remarkable competitiveness. On the demanding SWE-Bench Pro, Fugu Ultra achieved a score of 73.7, outperforming publicly accessible frontier models such as Opus 4.8 (69.2) and GPT-5.5 (58.6).
The system stands shoulder-to-shoulder with leading systems like Anthropic’s Fable 5 and Mythos Preview, matching or exceeding their performance thresholds across multi-turn reasoning and agentic tasks. Because Fugu secures this frontier-level performance through a decentralized pool of swappable components, global enterprises can leverage elite AI capabilities without the risk of being cut off by sudden changes in regional export control policies.
Early corporate testers and developers have noted substantial qualitative improvements over standard single-model deployments, particularly regarding thoroughness and the preservation of persona stability throughout extended working sessions.
