On September 11, Sakana AI released two new versions of Sakana Fugu, the system that bundles multiple AI models behind a single interface: Fugu Max, built around low cost, and Fugu Ultra v2, built around peak capability. Neither is a single giant model. Both route each task to a pool of available models depending on what the work requires. Fugu Max prices output tokens 40 to 60 percent below competing systems, while Fugu Ultra v2 beat models costing several times more on multiple benchmarks.

One architecture, split between cheap and strong

For the past decade, the AI industry has poured its resources into making models larger. Sakana AI takes a different position: the axis that matters in practice is two-dimensional, capability and cost, and firing up a multi-trillion-parameter model to run a simple lookup is not intelligence but waste. What the company is pushing is a system that decides not only how to solve a problem, but which machinery is the cheapest way to solve it.

The two releases are not separate products. They are the same orchestration architecture tuned for different missions. Fugu Max answers the question of what the best possible output is at the lowest possible cost. Fugu Ultra v2 answers the question of how far capability can be pushed on complex, multi-step work.

Fugu Max: more models to choose from, lower unit cost

Fugu Max significantly widens the set of models Fugu can route work to. It integrates open-weights models and specialized models at an unprecedented scale, including the NVIDIA Nemotron family through the company's collaboration with NVIDIA. By dynamically sending each task to the leanest model capable of solving it, the design aims for frontier-grade results at a fraction of the token spend.

Pricing is 2 USD (about 310 yen) per million input tokens and 6 USD (about 930 yen) per million output tokens. Sakana AI says that output price is 40 to 60 percent lower than Sonnet 5, GPT 5.6 Terra, and Kimi K3. On performance, Fugu Max records the best overall score on six benchmarks, including Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and SWEFish, the company's own internal benchmark. It also expands the cost-performance efficiency frontier on 7 of 10 benchmarks.

※1 USD = 155 JPY (as of September 11, 2026)

Fugu Ultra v2: reaching the top without holding the top models

Fugu Ultra v2 prioritizes output quality for complex multi-step reasoning, autonomous research, and full-stack software development. Its strength shows up on work that requires sustained reasoning over visual and structured data.

On Chartography, which tests visual reasoning and data interpretation, it scores 48.3, ahead of Opus 5 at 27.3 and Fable 5 at 29.5. On DeepSWE, a benchmark for real-world software engineering, it scores 74.3, beating models that cost three to five times more per token. Across eight benchmarks, it takes the best or joint-best score on five and places in the top two on seven.

The notable part is that these results come without Fable 5, Fable 5.1, or GPT-6 Astra in the routing pool. If frontier-grade output can be produced without holding a top-tier commercial model inside the system, then dependence on a single vendor and the risk of a sudden API shutdown both become easier to avoid. That structure is what Sakana AI means when it repeatedly stresses supply chain resilience.

Existing users can migrate with a one-line change

Both models are available from launch day through a standard OpenAI-compatible API. For anyone already running Fugu, switching to Max or Ultra v2 takes a single-line parameter change, with no migration work required. New architecture, same API.

Sakana AI also laid out the path Fugu has taken. In April, the beta showed that multi-agent orchestration can function as a unified foundation model. In June, general availability and Fugu Ultra v1 showed an orchestration layer matching closed frontier models on hard benchmarks. In July, Fugu-Cyber and a Claude Code interface extended it into specialized domains. In August, Sakana Chat carried it into daily consumer use while integration of NVIDIA open models began. This release continues that line.

How to read the numbers

That said, most of the scores presented were measured by Sakana AI itself, and one of the six benchmarks, SWEFish, is an internal benchmark the company built to reflect its own engineering challenges. Strong numbers on a metric designed around your own use cases are only to be expected, so how much independent third-party reproduction follows will be worth watching.

The routing approach itself also deserves a careful read. The cost advantage applies when the system correctly identifies which work a lean model can handle, and misjudgments add wasted round trips. The practical order is to measure what share of your own workload actually falls into the lighter category, then compare unit prices.

Summary

Sakana AI released Fugu Max, tuned for cost, and Fugu Ultra v2, tuned for capability, from the same orchestration architecture. Fugu Max charges 6 USD per million output tokens, 40 to 60 percent less than competitors, while Fugu Ultra v2 scores 48.3 on Chartography and 74.3 on DeepSWE. Reaching that level without top-tier commercial models in the routing pool points to a path forward other than simply making models bigger.