Fugu runs multiple models for the price of a single API

Sakana AI lance Fugu Max et Fugu Ultra v2, deux configurations multi-agents qui répartissent les tâches entre plusieurs modèles afin de privilégier le coût ou la qualité.

A single request may require difficult reasoning, document research, code generation, and verification. Sakana AI no longer wants to entrust all these steps to the same model. With Fugu Max and Fugu Ultra v2, the Japanese company aims to select several specialists, organize their work, and deliver their combined result through a single API.

Announced on September 11, 2026, the two services are not new foundation models trained from scratch. They are two configurations of Fugu, an orchestration layer capable of assigning roles to different models, sharing each participant’s output with the others, and changing the team’s composition according to the task.

For users, this organization remains largely invisible. An application sends its request through an OpenAI-compatible interface and receives a unified response. It does not have to call each provider separately, divide the work into stages, or reconcile multiple outputs.

Internally, however, the process can differ significantly from a traditional request. One model may propose a strategy, another may carry out part of the work, and a third may verify the result. Several exchanges may be required before the final response is produced. The composition and order of these interventions do not necessarily follow the same pattern for every request.

This approach draws on two earlier Sakana AI research projects. TRINITY uses a lightweight coordinator whose structure was obtained through automated evolution. Among other functions, it distributes thinking, execution, and verification roles across different models.

Conductor explores another method. The coordinator learns through reinforcement learning to formulate natural-language instructions for the other agents. It can also select itself as a participant, allowing recursive structures and additional compute when the difficulty warrants it.

Fugu turns this research into a service. Instead of asking customers to choose among several models from a static menu, Sakana AI attempts to make that decision during execution. A simple task can be routed to a less expensive resource, while a difficult problem may involve a larger team.

Fugu Max and Fugu Ultra v2 use the same general architecture, but they do not pursue exactly the same objective. Max seeks the best balance between quality and price. Ultra v2 prioritizes the final answer’s quality, even if that means involving more agents and increasing response time.

Sakana AI explains this difference through the concept of the Pareto frontier. A system sits on this frontier when no competitor can simultaneously provide better results at a lower cost. Improving quality at the same price, or reducing the price without sacrificing quality, moves that frontier outward.

Fugu Max is designed to extend it on the efficiency side. It uses the largest collection of models ever integrated into Fugu, including open and specialized systems from NVIDIA’s Nemotron family. The service dynamically searches for the combination it considers sufficient to handle each request.

This selection does not mean users can see the name of every participant in advance. Sakana AI does not publish the complete composition of Max’s pool or the exact path followed by each request. The phrase “open models” also does not guarantee that every component has downloadable weights.

The company credits Max with the best overall result on six evaluations: Terminal-Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench, and SWEFish. The last of these is an internal evaluation based on Sakana AI’s own programming work.

It also claims that Max expands the cost-performance frontier on seven of the ten tests shown. Depending on the competitor and evaluation selected, its results reportedly approach those of the strongest individual models at two to six times lower cost.

This promise rests partly on pricing of $2 per million input tokens and $6 per million output tokens. Cached input costs $0.25 per million tokens. Web search and page retrieval are billed separately at $0.007 per call.

Sakana AI specifically compares its $6 output rate with the prices of Sonnet 5, GPT-5.6 Terra, and Kimi K3, which it says are 40% to 60% higher. This comparison concerns listed unit prices. It is not enough to predict the cost of an entire task, since response length, the number of stages, and tool calls may vary from one system to another.

The company says it does not combine multiple model rates when several agents participate. Customers pay a single rate corresponding to the selected plan or, for the standard Fugu service, to the highest-tier model included in the pool. This rule simplifies billing, but it does not reveal how much compute is actually consumed behind each response.

Fugu Ultra v2 takes the opposite direction. It coordinates a deeper pool of specialists to maximize quality for autonomous research, multi-step reasoning, and full-stack software development. Sakana AI acknowledges that this organization increases response time.

Its most striking result comes from Chartography, an evaluation focused on interpreting charts and structured data. Ultra v2 scores 48.3 points, compared with 29.5 for Fable 5 and 27.3 for Opus 5 in the company’s table.

On DeepSWE, which covers software development tasks, its score reaches 74.3. Sakana AI says it therefore surpasses models that cost three to five times more per token. Ultra v2 also ranks first or joint first on five of the eight selected tests: GDP.pdf, Chartography, DeepSWE, Toolathon, and SWEFish. It reportedly places among the top two systems on seven evaluations.

These results do not come from an individual model. They measure a complete system comprising the coordinator, the various agents, their exchanges, the available tools, and the resolution budget. A comparison with a model called directly therefore evaluates two different forms of computation.

The selection of reference models also matters. Sakana AI emphasizes that Fable 5, Fable 5.1, and GPT-6 Astra are not included in