Editorial illustration for Sakana AI's New Fugu Models Add Nemotron for Cheaper Multi-Agent AI
Sakana AI's Fugu Adds Nvidia Nemotron for Cheaper AI
Sakana AI added Nvidia's Nemotron models to the mix behind its Fugu system this week, part of a broader release that splits the product into two versions: Fugu Max and Fugu Ultra v2. Neither is a single foundation model. Fugu works as a learned orchestrator, routing each incoming query across a pool of underlying models through one API, and deciding on the fly which combination of that pool can actually handle the task.
The two new releases split the mission in half. Fugu Max is built for cost, aiming to get the best output per dollar by avoiding overkill on simple requests. Fugu Ultra v2 is built for ceiling performance, meant for harder multi-step work where the price tag matters less than getting the answer right. Both are live now through Sakana's OpenAI-compatible API, hosted rather than open-weight, and unavailable in the EU/EEA.
The pace here has been quick. Fugu went into beta in April, hit general availability in June, then picked up a Fugu-Cyber variant and a Claude Code interface in July. Sakana's own framing of why any of this matters comes down to a specific tradeoff the company keeps coming back to.
Sakana’s main message is frontier output without dependence on any 1 proprietary model. The research team states that this reduces exposure to vendor lock-in, API revocations, and sudden service cutoffs.
Why this matters
For teams paying per-token bills across multi-agent pipelines, Fugu Max's routing logic is the interesting part, not the Nemotron addition itself. Sending each task to the cheapest model that can still solve it is a real cost lever if the routing decisions hold up under production load, not just benchmark conditions. Nvidia's involvement gives Sakana AI a credibility boost and a wider model pool to draw from, but it also means Fugu's economics now depend partly on Nemotron's pricing and availability, not just Sakana's own infrastructure.
The bigger question for developers evaluating this: how transparent is the routing decision itself? An orchestrator that silently downgrades a task to a weaker model to save money is only useful if you can audit that choice, or override it when accuracy matters more than cost. Fugu Ultra v2 answers the inverse case, complex multi-step work where you want the strongest available model regardless of price.
Having both tiers live on one OpenAI-compatible API is a genuine convenience. Whether the routing is smart enough to trust with production traffic is the thing to watch once real usage data comes in.
Related Reading
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
From Archives Meta’s AI App Muse Is Second in U.S. Behind ChatGPT's Launch
From Archives OpenAI's Agents API Enters Public Beta, Creating Agents With One Call
Common Questions Answered
How does Sakana AI's Fugu system route queries across multiple models?
Fugu works as a learned orchestrator that routes each incoming query across a pool of underlying models through a single API, deciding on the fly which combination of models can best handle the task. This approach allows the system to optimize for either cost or performance depending on the specific use case and requirements.
What is the difference between Fugu Max and Fugu Ultra v2?
Fugu Max is built for cost optimization, routing tasks to the cheapest models capable of solving them, while Fugu Ultra v2 is designed for maximum performance and quality output. The split allows teams to choose the version that best aligns with their priorities, whether that's reducing token-based billing costs or achieving frontier-level results.
Why does Sakana AI emphasize avoiding dependence on a single proprietary model?
By not relying on any single foundation model, Sakana AI reduces exposure to vendor lock-in, API revocations, and sudden service cutoffs that could disrupt operations. This multi-model approach provides teams with greater resilience and flexibility in their AI infrastructure, protecting them from the risks associated with depending on one provider's service.
What role does Nvidia's Nemotron play in the new Fugu releases?
Nvidia's Nemotron models were added to the pool of underlying models that power Fugu's orchestration system, expanding the available options for routing decisions. Nvidia's involvement provides Sakana AI with credibility and access to a wider model pool, though it also means Fugu's economics now partly depend on Nemotron's pricing structure.
How can Fugu Max's routing logic help teams reduce their multi-agent pipeline costs?
Fugu Max's routing logic sends each task to the cheapest model that can still solve it effectively, creating a real cost lever for teams paying per-token bills across multi-agent pipelines. This optimization is particularly valuable if the routing decisions maintain accuracy and performance under production load conditions, not just in benchmark testing scenarios.
Further Reading
- Sakana AI Teams With NVIDIA to Advance Open Model Innovation - Sakana AI
- Japan's Enterprises and Startups Build Industry-Specialized AI With NVIDIA Nemotron Open Models - NVIDIA News
- Sakana AI's orchestrator adds Nvidia Nemotron to prove collective intelligence can rival single frontier models - The Decoder
- Sakana AI Wires NVIDIA Nemotron Into Fugu, Beating Claude and GPT-5.5 on Coding - AlphaSignal
- Introducing Fugu Max and Fugu Ultra v2: Orchestrating the Next Generation of Multi-Agent AI - Sakana AI
