The Japanese AI company Sakana AI introduced two new orchestration models, Fugu Max and Fugu Ultra v2, on September 10, 2026, which automatically distribute requests to a network of open and specialized systems. According to Sakana AI, Fugu Max costs up to 60 percent less than comparable offerings from Anthropic, OpenAI, and Moonshot AI. For now, the service remains blocked in the EU and the European Economic Area.
The orchestrator picks the cheapest model for each task
Fugu is not a single language model but an orchestrator: a trained coordinator receives a request, distributes sub-tasks to multiple models, and combines the responses into a result. The technology is based on two methods from Sakana AI’s own research: TRINITY assigns recurring roles such as thinker, executor, and verifier to individual work steps, while another system trained with reinforcement learning called Conductor develops suitable coordination strategies in natural language.
Fugu Max costs 2 US dollars per million input tokens and 6 US dollars per million output tokens and leads the rankings in six benchmarks, including Terminal Bench 2.1, GPQA Diamond, and the agent test SWEFish. One million output tokens correspond to about 750,000 words – roughly equivalent to 1,500 printed book pages. These figures come from Sakana AI and are independently unverified. For the model pool, the company collaborates with Nvidia and incorporates its Nemotron model family. The more powerful Fugu Ultra v2 pursues a different goal: it aims to achieve the highest quality on demanding, multi-step tasks and specifically excludes the models Fable 5, Fable 5.1, and GPT-6 Astra from the pool to reduce dependence on individual providers. Both new variants run exclusively through a hosted, OpenAI-compatible API. Sakana AI does not offer its own model weights for self-hosting. Existing customers are expected to switch with a single line of code change.
Sakana AI frames the pricing as a challenge to market leaders
According to industry outlets such as Yahoo Finance, Sakana AI shifts competition from individual model prices to the economics of orchestration itself with this pricing model. Because Fugu routes tasks to interchangeable, partly open models instead of being tied to a single proprietary system, the company partially decouples itself from the pricing decisions of individual providers.
The move joins a string of price cuts and discount campaigns across the industry. OpenAI cut the prices of its GPT-5.6 models by up to 80 percent over the summer, and Meta opened its own AI agent price war with Muse Spark 1.1. At the same time, IT buyers have already been struggling with an increasingly overwhelming flood of models from major providers – an additional orchestration model like Fugu at least promises to automate model selection under the hood instead of leaving it to customers themselves. It is not yet known whether large providers such as Anthropic or OpenAI feel pressured by cheaper orchestration layers. Public reactions from either company to the Fugu launch are not yet available.
What matters now is whether the orchestration approach establishes itself as an independent business model, or whether large providers simply fold similar routing layers into their own products and make independent middle layers like Fugu unnecessary. It also remains open when, or if, Sakana AI will open the service to customers in the EU and the European Economic Area. The company has not named a date for that.
