Tokyo-based Sakana AI on September 11, 2026 released Fugu Ultra v2.0, the newest quality-tier flagship in its Fugu family of multi-agent orchestration models. Delivered through a single OpenAI-compatible endpoint, Fugu Ultra v2.0 dynamically routes each request across a proprietary pool of frontier open-weights and specialized models, ships a 1,000,000-token context window, and lands top or joint-top scores on five of eight of Sakana's headline agentic benchmarks.
Multi-agent, delivered as one model
Sakana positions Fugu not as a single monolithic model but as a "multi-agent system delivered as one model": Fugu Ultra learns how to assemble and orchestrate expert agents from a swappable pool rather than following hand-designed workflows. Two ICLR 2026 papers, TRINITY: An Evolved LLM Coordinator and Learning to Orchestrate Agents in Natural Language with the Conductor, sit under the hood, describing how the system trains coordinators to assign Thinker, Worker and Verifier roles across coding, math and reasoning tasks.
Benchmarks and pricing
On agentic evaluations, Sakana reports that Fugu Ultra v2.0 places in the top two on seven of eight benchmarks including GDP.pdf, Chartography, DeepSWE, Toolathon and Sakana's internal SWEFish coding benchmark. Fugu Ultra is priced at $5 per 1M input tokens and $30 per 1M output tokens, with $10/$45 rates for contexts above 272K tokens; cached input starts at $0.50 per 1M. Alongside Ultra, Sakana positioned Fugu Max at $2 input and $6 output regardless of context length and unveiled Fugu Cyber, a security-specialized variant that Sakana says reaches 86.9% on CyberGym and 72.1% on CTI-REALM.
A hedge against single-vendor risk
Sakana explicitly frames Fugu Ultra as a hedge against single-vendor risk: because Ultra's output is produced by orchestrating a pool of open-weights and specialized models rather than a single proprietary frontier model, the company argues customers get frontier-tier performance without lock-in, sudden API revocations or geopolitical service cutoffs. The model is available worldwide through Sakana's API, on OpenRouter and via the Vercel AI Gateway, but is not offered inside the EU or EEA pending GDPR compliance work. The launch tracks the wider agentic-AI push already covered on The Robotics Media, including DeepSeek's V4.1-Flash 1M-context release, OpenAI's Agents API public beta, and Salesforce's Trusted Enterprise AI control plane.
Reporting based on coverage from Sakana AI, OpenRouter, DataNorth, LLM Reference and BenchLM.