Sakana Fugu Max and Ultra v2: What an AI Orchestrator Does

Sakana AI Ships Two New Fugu Orchestrators

Tokyo-based Sakana AI released Fugu Max and Fugu Ultra v2 on September 11, the latest step for a product line that takes an unusual route to frontier performance. Fugu is not a single foundation model. It is a learned orchestrator: a system trained to read a request, assemble an agent scaffold on the fly, and send each piece of the work to whichever model in its pool is best suited to it, including further copies of itself.

• Two versions, one architecture: Fugu Max targets the best output per dollar, while Fugu Ultra v2 targets peak capability on hard, multi-step tasks
• Fugu Max pricing: $2 per million input tokens and $6 per million output tokens
• Headline result (Sakana's figure): Ultra v2 scores 48.3 on the Chartography benchmark versus 27.3 for Claude Opus 5
• Availability: live now through Sakana's OpenAI-compatible API, where existing users switch with a one-line parameter change

How Is an AI Orchestrator Different From a Single Model?

A conventional large language model answers everything with the same weights, whether the task is formatting a date or refactoring a codebase. An orchestrator treats models as a team. Simple sub-tasks go to small, cheap specialists; hard reasoning goes to stronger models; and a verification pass can be handed to a different model entirely so errors get caught before the answer returns.

Sakana's pool mixes open-weight and specialized models, including NVIDIA Nemotron models through the company's collaboration with NVIDIA. Notably, Sakana says Ultra v2 reaches its results without Fable 5, Fable 5.1 or GPT-6 Astra in its agent pool, which is a pointed claim about how far coordination alone can carry a system.

According to MarkTechPost, the approach builds on two of Sakana's ICLR 2026 papers: TRINITY, which assigns agents Thinker, Worker or Verifier roles, and The Conductor, which uses reinforcement learning to discover coordination strategies. The company also draws on fine-tuning and evolutionary methods, a Sakana hallmark.

What Sakana Reports on Benchmarks

Every comparison below is Sakana's own, and no independent evaluation has been published yet.

• Fugu Max posts the best overall score on six benchmarks: Terminal Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench and SWEFish. Sakana says its output price is 40% to 60% lower than Sonnet 5, GPT 5.6 Terra and Kimi K3
• Fugu Ultra v2 is best or joint-best on five of eight benchmarks and in the top two on seven of eight. Beyond Chartography, where Fable 5 scored 29.5, Sakana reports 74.3 on DeepSWE

The pattern is the story. If a well-trained router over a pool of mostly open models can compete with single frontier systems on agentic coding and document work, the economics of multi-agent AI shift toward whoever coordinates best, not only whoever trains the largest model.

What Should Developers Know Before Trying Fugu?

Integration is deliberately boring: Fugu speaks the OpenAI-compatible API format, so existing client code needs a new base URL and model name rather than a rewrite. Ultra v2 is also listed on OpenRouter. There are no open weights for self-hosting, and MarkTechPost reports the service is not available in the EU or EEA, with an Ultra v2 training cutoff of August 28, 2026.

The release caps a fast year. Fugu entered beta in April, reached general availability in June, added the Fugu-Cyber specialization in July, and gained its NVIDIA partnership in August. Routing is a crowded idea right now, from Cursor's automatic model router to the gateway layer behind Stripe's OpenRouter deal. Sakana's bet is that the router itself should be a trained model. Follow the rest of the week's launches in our AI coverage.

Sources: Sakana AI — September 11, 2026; MarkTechPost — September 10, 2026; OpenRouter: Fugu Ultra v2 listing — September 2026.