What Fugu Ultra Is and Why It Matters
Fugu Ultra is a multi-agent AI orchestration system that coordinates a pool of language models through one controller model, delegating subtasks, merging outputs, and routing work so that users experience a single, higher-performing assistant instead of many separate tools. Sakana AI positions Fugu Ultra as a frontier-level orchestration model, competing with leading systems by focusing on AI model coordination rather than training one giant model from scratch. It runs through a single OpenAI compatible API, so developers can plug it in much like any other large model, while Fugu Ultra quietly calls specialized agents in the background. By targeting tasks in engineering, science, research, cybersecurity, and data analysis, it aims to give smaller teams access to competitive AI systems without accepting single-vendor risk or depending on export-restricted infrastructure.
How Multi-Agent AI Orchestration Works in Fugu
Sakana Fugu and Fugu Ultra center on the idea that a coordinated team of models can outperform any one model alone. Instead of wiring fixed workflows, Fugu learns how to assemble agents dynamically from its pool, routing subtasks and deciding when to call Thinker, Worker, or Verifier roles. According to Sakana’s description of its TRINITY and Conductor research, the coordinator learns natural-language strategies for prompting and combining agents, rather than relying on hand-designed pipelines. This approach means Fugu can call instances of itself and other expert LLMs, then synthesize their answers into one response. The result is multi-agent AI orchestration delivered as a single endpoint, sparing developers from juggling multiple providers. As new models join the pool, the orchestration layer can improve without users changing their integrations or workflows.
Benchmarks: Matching Frontier Labs Without a Monolithic Model
Sakana’s benchmark scores suggest that orchestration can match or surpass top standalone models on several demanding tests. The company reports that Fugu Ultra reaches 73.7 on SWE-Bench Pro, ahead of Claude Opus 4.8’s 69.2 and GPT-5.5’s 58.6, and scores 93.2 on LiveCodeBench compared with Gemini 3.1 Pro’s 88.5. On Humanity’s Last Exam, Fugu Ultra scores 50.0, essentially tied with Opus 4.8’s 49.8. Sakana also highlights qualitative results: in an AutoResearch run with 123 training experiments on a single H100 GPU, Fugu Ultra delivered the best mean validation score among tested models, and an industry researcher cut a multi-day patent landscape analysis down to a few hours. These results support Sakana’s claim that careful AI model coordination can rival the performance of larger, monolithic frontier systems.
APIs, Pricing, and Vendor Flexibility for Enterprises
Fugu Ultra is exposed through an OpenAI compatible API, making it easy for development teams to drop into existing stacks that already speak that interface. Sakana offers two tiers: Fugu for everyday coding, review, and chatbot use, and Fugu Ultra for long workflows like Kaggle competitions, paper reproduction, cybersecurity assessments, and patent or literature investigations. When only one agent is active, customers pay the base rate for that underlying model; when multiple agents coordinate, they pay a single rate tied to the top-tier model involved. Fugu Ultra itself is priced at USD 5 (approx. RM23) per million input tokens and USD 30 (approx. RM138) per million output tokens, with subscription plans at USD 20 (approx. RM92), USD 100 (approx. RM460), and USD 200 (approx. RM920) per month. Enterprises can choose which providers participate in the pool, gaining frontier-level performance without locking into one vendor.
From Compute Arms Race to Orchestration Advantage
Fugu Ultra signals a shift in AI competition from building ever-larger single models to extracting more value through orchestration efficiency. Sakana’s founders have long argued for alternatives to brute-force scaling, and Fugu embodies that stance: it treats existing models as components in a coordinated system rather than endpoints. The system’s emphasis on sustained persona consistency, thoroughness in long workflows, and resilience against single-vendor dependence makes it attractive to teams that want competitive AI systems but cannot or will not match frontier labs on compute budgets. For smaller startups, Sakana Fugu Ultra is a proof-of-concept that coordination strategies can open a new lane in the market. Instead of chasing the biggest model, they can compete by building smarter conductors that turn many good models into one strong, coherent assistant.
