Sakana’s Fugu AI system rivals top models without vendor lock-in

AI infrastructure is undergoing a quiet revolution—and Sakana AI just fired a major shot across the bow.

Last week, the Tokyo-based AI startup introduced Fugu, a multi-agent orchestration system designed to deliver frontier-level performance through a single, OpenAI-compatible API. Unlike traditional monolithic models, Fugu dynamically routes user queries to a rotating pool of specialized AI agents, reducing reliance on any single provider and mitigating risks tied to sudden regulatory changes.

This approach comes at a pivotal moment. Anthropic’s recent decision to restrict public access to its most advanced models, including Claude Fable 5 and Mythos 5, underscored the fragility of vendor-dependent AI deployments. Sakana’s response? A system that doesn’t just adapt to these constraints—it actively bypasses them.

A smarter, more resilient alternative to single-model AI

Fugu operates like a high-stakes project manager. Instead of trying to solve complex problems alone, it deconstructs tasks, assigns sub-problems to expert models, validates their outputs, and synthesizes the final result. This multi-agent strategy is grounded in two of Sakana’s 2026 research papers: TRINITY and Conductor, which explore autonomous coordination strategies rather than rigid, hand-crafted workflows.

"Fugu is itself an LLM trained to invoke various LLMs in an agent pool, including recursive calls to itself," explained the Sakana AI team in their technical announcement. By acting as a broker rather than a standalone model, it eliminates the single-point-of-failure risk that plagues systems dependent on a single foundation model.

Sakana offers two service tiers to match different operational needs:

• Fugu: Optimized for speed and low latency, ideal for interactive chatbots and real-time coding environments like Codex.
• Fugu Ultra: Engineered for high-stakes use cases such as AI research, cybersecurity analysis, and multi-step patent investigations. It leverages a deeper pool of specialized agents and reportedly matches or exceeds the performance of leading monolithic models on rigorous benchmarks.

Pricing reflects this specialization. The standard Fugu model uses a dynamic pay-as-you-go model based on activated agents, while Fugu Ultra enforces a fixed rate starting at $5 per million input tokens and $30 per million output tokens.

Benchmark results that challenge the status quo

Sakana’s claims aren’t theoretical. On LiveCodeBench, an open benchmark tracking real-time coding performance, Fugu’s standard and Ultra variants outperform Anthropic’s Claude Fable 5:

• Fugu Ultra: 93.2
• Fugu (standard): 92.9
• Claude Fable 5: 89.8

On GPQA-Diamond, a rigorous test of graduate-level reasoning in biology, physics, and chemistry, both Fugu variants achieve 95.5, surpassing the prior Claude Mythos Preview score of 94.6.

These results suggest that orchestration-driven AI isn’t just a theoretical advantage—it’s a practical one. By distributing workloads across multiple providers, Fugu builds redundancy into the stack. If one provider faces an outage or regulatory block, the system reroutes automatically, maintaining uptime and performance.

Control, compliance, and strategic trade-offs

Fugu is a proprietary API service, not an open-source framework. Its core advantage lies in proprietary routing logic—the exact models selected for each query remain hidden from users, ensuring competitive differentiation.

For enterprises sensitive to data residency and privacy, Sakana provides granular controls. Users can explicitly exclude specific models or providers from their routing pool, enforcing strict compliance with corporate policies. Additionally, they can opt out of prompt data collection for future training—a critical feature for organizations handling sensitive or proprietary information.

Geographically, Fugu is currently unavailable within the European Union and European Economic Area as Sakana navigates alignment with regional data protection frameworks. The company has not indicated a timeline for EU-wide availability.

The future of AI infrastructure is distributed

David Ha, Sakana’s CEO and co-founder—formerly of Google Brain—argues that orchestration models like Fugu represent the next frontier beyond ever-larger foundation models. "Relying on a single company’s model for national infrastructure is a massive risk," he wrote in a recent post. "Collective intelligence is the practical hedge against this concentration of power."

As geopolitical and regulatory pressures intensify, the shift away from monolithic AI models is accelerating. Systems that can dynamically adapt, reroute, and scale across providers are no longer optional—they’re essential. Fugu isn’t just another API; it’s a blueprint for a more resilient, vendor-agnostic AI future.