Sakana Fugu offers a new way to coordinate frontier artificial intelligence models without relying on a single vendor.
Developed by Tokyo-based startup Sakana AI, this system acts as an intelligent router. It directs user queries to a pool of top-tier models like GPT-5.5, Claude Opus, and Gemini 3.1 Pro. Instead of managing multiple APIs, developers can access collective intelligence through a single endpoint. This approach helps businesses maintain operational stability even when specific providers change their access terms.
How Sakana Fugu Coordinates Expert Models
The core technology relies on learned orchestration methods. According to the official Sakana Fugu release notes, the system uses two key frameworks called Trinity and Conductor. Trinity assigns specific roles such as Thinker, Worker, or Verifier to different models during a task. Meanwhile, Conductor uses reinforcement learning to discover the best natural-language coordination strategies. This allows the platform to solve simple tasks directly or assemble a specialized team of models for complex, multi-step problems.
Bypassing Single-Vendor Risks
Enterprise clients face growing challenges with sudden policy shifts and export controls. For instance, recent restrictions on Anthropic’s Fable and Mythos models disrupted several global workflows. Sakana AI CEO David Ha highlighted that this platform serves as a critical hedge. If one provider restricts access, the system automatically routes tasks to other available models. This ensures that businesses operating in the economy sector do not experience sudden service interruptions.
Performance Benchmarks and Real-World Testing
The platform launches with two versions: a standard model and Fugu Ultra. The Ultra version is designed for maximum accuracy on demanding engineering and scientific tasks. On the SWE-Bench Pro benchmark, Fugu Ultra achieved a score of 73.7, outperforming many individual models. A report by VentureBeat notes that the orchestrator often beats the individual models it coordinates. However, some independent testers have reported a gap between benchmark claims and real-world performance, highlighting the need for thorough testing.
Pricing and Global Availability
Subscriptions for the service are structured at three tiers: $20, $100, and $200 per month. For high-volume users, pay-as-you-go pricing for the Ultra model is set at $5 for input and $30 for output per million tokens. The service is available globally, though it is temporarily blocked in the EU and EEA while the company works toward GDPR compliance. Developers looking to integrate Sakana Fugu can access it via an OpenAI-compatible API, making it easy to add to existing apps and software workflows.
