The Sakana AI Fugu Max and Fugu Ultra v2 have arrived (September 11, 2026). Tokyo-based startup Sakana AI has launched two artificial intelligence orchestration systems: instead of relying on a single giant model, a "maestro" selects, combines, and reviews multiple models tailored for each specific task.
The announcement was shared on the company's official blog and highlighted by ITmedia NEWS in Japan. Both products are already available via an OpenAI-compatible API. For developers following the recent wave of agent APIs—such as OpenAI's Agents API—this presents a fresh alternative: deploying a team of models rather than just one.
Image credit: Sakana AI / David Ha — Source: Official announcement on X
What happened?
Orchestration, in this context, means the system receives a prompt, determines which model is cost-effective and capable enough to handle it, potentially calls auxiliary models to cross-check the output, and only then returns the final answer. This counters the traditional "one massive model for everything" approach.
The Sakana AI Fugu Max is engineered for cost-efficiency. According to the company, it charges $2 per million input tokens and $6 per million output tokens. Sakana claims its output pricing runs 40% to 60% lower than Claude Sonnet 5, GPT-5.6 Terra, and Kimi K3. The model pool features both open-weights and specialized models, including NVIDIA's Nemotron family.
Meanwhile, the Fugu Ultra v2 targets maximum performance. The company states it ranked first or tied for first in five out of eight internal and public benchmarks, including Chartography (48.3 versus Opus 5's 27.3) and DeepSWE (74.3). Notably, ITmedia pointed out that Ultra v2's pool does not include GPT-6 Astra, Claude Fable 5, or Fable 5.1; Sakana asserts it surpasses these models without using them under the hood.
What does this mean?
For developers, the pitch is a single API with two flavors. Existing Fugu users can switch versions by simply tweaking a parameter. The Ultra v2 costs $5 / $30 per million tokens (input/output), with input caching priced at $0.50. Context windows exceeding 272k tokens fall into a premium tier, according to launch reports. Third-party catalogs also list a 1-million-token context window for Ultra v2.
Sakana's core thesis revolves around the Pareto frontier: being the strongest model isn't enough; what matters is the optimal balance between performance and price. The Sakana AI Fugu Max shifts this point toward cost savings, while the Ultra v2 leans toward top-tier quality. In a market where Chinese models like DeepSeek-V4.1-Flash are also driving prices down, orchestration offers another viable route to cut expenses.
What it doesn't mean yet
Benchmark metrics currently rely on Sakana's own testing. Switching models mid-conversation can consume extra orchestration tokens and reduce cache reuse efficiency. Reports from the first generation of Fugu indicated potential latency and rapid credit depletion. The company did not disclose cache hit rates or the average total cost per task for the Max in this announcement.
Context
Founded in 2023 in Tokyo by David Ha and Llion Jones, Sakana AI pursues a path distinct from American tech giants: collective intelligence inspired by fish schools (the name means "fish" in Japanese). Fugu exited beta in April, received an Ultra version in June, integrated with Claude Code in July, and formed a partnership with NVIDIA in August. The technical report cited in the announcement is available on arXiv.
In practice, Sakana AI Fugu Max appeals to those seeking cutting-edge quality without paying frontier-model prices for every single call. Ultra v2 caters to users prioritizing long-form tasks—such as coding, data charting, and deep research—who are willing to pay more per response.
Sources: Sakana AI, ITmedia NEWS, and AI Weekly coverage.
By GeekikiBot
