Sakana Launches Fugu Orchestrator to Route Around Frontier AI Export Blocks
TOKYO — Shifting the enterprise AI focus from monolithic models to collective intelligence, Tokyo-based startup Sakana AI has launched Fugu. The multi-agent orchestration system offers developers a single, OpenAI-compatible API that matches frontier performance while providing a resilient hedge against single-vendor lock-in and geopolitical export controls.
The system emerges as a vital infrastructure alternative following Anthropic’s sudden decision on June 12 to pull its flagship models, Claude Mythos 5 and Claude Fable 5, from global public access to comply with a U.S. government export restriction order.
Collective Swarm Intelligence Over Brute Force
Founded by former Google Brain researcher Llion Jones and Stability AI's David Ha, Sakana AI rejects the hyper-fixation on ever-larger single models, favoring systems built around biomimicry and evolutionary "swarm" intelligence.
Fugu operates as an automated general contractor:
- Dynamic Task Breakdown: When handed a complex query, Fugu uses an underlying LLM layer to slice the problem into sub-tasks.
- Recursive Delegation: Micro-tasks are dynamically dispatched across a fluid, swappable pool of third-party expert models.
- Synthesis & Verification: The system automatically cross-checks code or reasoning steps recursively before compiling the final user response.
To protect corporate compliance, developers can explicitly opt specific model families or providers out of their active routing pipeline. The entire routing and synthesis process is black-boxed from the user by design to protect Sakana's core collaboration logic. Geographically, the service is temporarily restricted within the EU and EEA while the team aligns its black-box routing paths with GDPR.
Performance Trade-Offs & Multi-Agent Billing Realities
Fugu is offered in a high-speed standard tier and a flagship Fugu Ultra (fugu-ultra-20260615) tier meant for high-stakes tasks like AI research and cybersecurity analysis. On pay-as-you-go plans, standard Fugu charges the single rate of the highest-tier underlying model used. Fugu Ultra employs a premium, fixed pricing matrix: $5.00 per 1M input tokens, $30.00 per 1M output tokens, and $0.50 per 1M cached tokens (with rates scaling up to $10 / $45 for workloads above 272K context).
Enterprise Cost Warning: Unlike traditional wrappers, Fugu Ultra passes the background token costs of internal agent-to-agent delegation, code verification, and recursive reasoning loops straight to the final user bill.
Benchmark Highlights
Collective intelligence proves highly potent on complex, multi-step agentic tasks, though it can still lag behind single-domain monolithic beasts when direct access is available:
- Coding Performance: On LiveCodeBench, Fugu Ultra hit 93.2, outperforming Anthropic's restricted Claude Fable 5 (89.8). It also notched a 73.7 on SWE-Bench Pro, comfortably beating OpenAI's GPT-5.5 (58.6).
- Scientific Reasoning: Fugu Ultra scored 95.5 on the graduate-level GPQA-D (Diamond) exam, beating the prior Claude Mythos Preview (94.6).
- The Monolithic Edge: Despite these wins, Fugu Ultra still trails the restricted Fable 5 on pure software engineering execution (73.7 vs 80.0) and drops slightly behind GPT-5.5 on long-context recall benchmarks (93.6 vs 94.8).
Source & References
- Primary Coverage: No Claude Fable 5? No problem: Sakana achieves frontier performance with new Fugu multi-model, auto synthesis system — VentureBeat
Struggling to balance performance caps against single-vendor reliance? Head over to the ChooseAIModel Directory to track live enterprise pricing, context parameters, and latency metrics across hundreds of active foundational options. To optimize your multi-agent architecture, use our free Cost Simulator to instantly stress-test and forecast your production token spend before your next deployment.