Sakana AI Launches Fugu Max and Fugu Ultra v2 for Cheaper, Stronger Multi-Agent Orchestration
MarkTechPost Asif Razzaq ● Covered by 2 sources
Sakana’s new Fugu models route jobs across other AI models to cut cost or push quality. That’s the point: one API, no self-hosting, and no EU/EEA access.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Sakana AI has pushed out two new Fugu models, and the pitch is refreshingly practical: use the right machine for the job, not the fanciest one in the room. Fugu Max is aimed at getting the best result for the least money. Fugu Ultra v2 is aimed at squeezing out the highest capability on difficult, multi-step work. Both are live now through Sakana’s OpenAI-compatible API.
Fugu is not a single foundation model pretending to do everything. It is a learned orchestrator that takes a request, builds an agentic scaffold, and routes the work across a pool of other models behind one API. Sakana says the system is trained with large-scale fine-tuning, evolutionary algorithms, and reinforcement learning, and that the release builds on two ICLR 2026 papers: TRINITY and The Conductor. The company’s argument is simple enough. Real workloads sit on a cost-capability tradeoff, and a system should spend as little as possible while still solving the task.
Fugu Max is the thriftier of the two. Sakana widened the model pool it can call on, including open-weights and specialized models, plus the NVIDIA Nemotron family through a collaboration with NVIDIA. The company says it routes each task to the leanest model that can handle it. Pricing is listed at $2 per 1 million input tokens and $6 per 1 million output tokens. Sakana says that puts output prices 40% to 60% below Sonnet 5, GPT 5.6 Terra, and Kimi K3, while also claiming the best overall score on six benchmarks: Terminal Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench, and SWEFish.
Fugu Ultra v2 is the more ambitious sibling. Sakana says it is built for complex reasoning, autonomous research, and full-stack software development, with especially strong gains on visual and structured data. On Chartography, the company reports 48.3, compared with 27.3 for Opus 5 and 29.5 for Fable 5. On DeepSWE, it reports 74.3, ahead of models priced 3x to 5x higher per token. It is also said to be best or joint-best on five of eight benchmarks and in the top two on seven of eight. Fugu Max and Fugu Ultra v2 share the same orchestration architecture, but only these two targets differ.
There are also some hard limits. Sakana says there are no open weights to self-host, and the service is not available in the EU/EEA. The company is leaning hard into the anti-lock-in message at the same time it keeps the keys to the system. That’s a very modern kind of openness.
My take — AI-written commentary, not fact-checked reporting
This is the right bet, and also the least romantic one: orchestration beats worshipping a single giant model. Closed systems keep proving they’re useful when they’re disciplined, and open-weight fans can keep arguing in the comments while the cost tables do the talking. The EU/EEA exclusion, though, is the old AI story in one line: sell the future, then hide the door.
Read more about this at: MarkTechPost
