Sakana AI has released Fugu Max and Fugu Ultra v2, two new models designed to route work across a pool of other systems via a single API. Fugu is not a foundation model but an orchestrator that manages tasks behind the scenes. The new versions tune this architecture for two specific goals: Fugu Max aims for the best output per dollar, while Fugu Ultra v2 targets the highest capability on difficult, multi-step tasks.
Both models are available today as hosted APIs compatible with OpenAI standards. There are no open weights for self-hosting, and the service is not offered in the EU or EEA.
Why Sakana Frames This as a Two-Axis Problem
Sakana’s argument is direct. Real workloads are judged on capability and cost together. Sending a simple data lookup to a multi-trillion-parameter model wastes money. A better system picks the cheapest machinery that can still solve the task.
The team describes this with the Pareto frontier. On that frontier, gaining quality costs more, and cutting cost loses quality. Fugu Max and Fugu Ultra v2 share one core orchestration architecture. Only the optimisation target differs.
The release follows a fast cadence. Fugu entered beta in April, reached general availability in June, and added Fugu-Cyber and a Claude Code interface in July.
How Fugu Orchestration Works
The Sakana Fugu Technical Report describes Fugu models as language models in their own right. They read a query and build an agentic scaffold for it on the fly. Training combines large-scale fine-tuning, evolutionary algorithms, and reinforcement learning.
The system builds on two ICLR 2026 papers. TRINITY uses a lightweight evolved coordinator that assigns Thinker, Worker, or Verifier roles across turns. The Conductor is trained with reinforcement learning to discover natural-language coordination strategies and focused prompts.
Fugu Max: More Models, Less Cost
Fugu Max widens the pool of models Fugu can orchestrate. It adds a large set of open-weights and specialised models. That includes the NVIDIA Nemotron family, through Sakana’s collaboration with NVIDIA. Fugu Max routes each task to the leanest model capable of solving it.
Sakana team reports the following:
- Pricing: $2 per 1M input tokens and $6 per 1M output tokens.
- Output price: 40% to 60% lower than Sonnet 5, GPT 5.6 Terra, and Kimi K3.
- Performance: Best overall score on 6 benchmarks: Terminal Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench, and SWEFish.
- Efficiency: Expands the cost-performance Pareto frontier on 7 of 10 benchmarks.
Sakana places Fugu Max within striking distance of elite models at 2x to 6x lower cost. SWEFish is an internal Sakana benchmark built from its own coding challenges. Treat that result as a vendor signal.
Fugu Ultra v2: Raising the Ceiling
Fugu Ultra v2 targets complex reasoning, autonomous research, and full-stack software development. Its largest gains appear on sustained reasoning over visual and structured data.
- Chartography (visual reasoning and data interpretation): 48.3, versus 27.3 for Opus 5 and 29.5 for Fable 5.
- DeepSWE (real-world software engineering): 74.3, ahead of models priced 3x to 5x higher per token.
- Breadth: Best or joint-best on 5 of 8 benchmarks: GDP.pdf, Chartography, SWEFish, DeepSWE, and Toolathon.
- Consistency: Top 2 on 7 of 8 benchmarks.
Fable 5, Fable 5.1, and GPT-6-Astra are not in Fugu Ultra v2’s agent pool. The model’s training cutoff is August 28, 2026. Sakana’s main message is frontier output without dependence on any single proprietary model. The research team states that this reduces exposure to vendor lock-in, API revocations, and sudden service cutoffs.
What it means
For developers building software agents, this shift changes the economics of scaling. Previously, teams had to choose between expensive, high-end models or cheaper, less capable ones. Now, the orchestrator handles the routing. You can run complex workflows where the system automatically delegates simple steps to cheaper models and reserves expensive power only for tasks that genuinely require it. This removes the need to manually tune which model handles which part of a pipeline.
