Sakana Fugu Max and Fugu Ultra v2 Route Tasks to Models

A model whose job is choosing models

Sakana's Fugu technical report describes Fugu models as language models trained to understand a user's query and devise, on the fly, an agentic scaffold to solve it, handing the work to other models in a pool. Sakana released Fugu in beta in April 2026 and made it generally available with Fugu Ultra v1 in June. [3][1]

The September release is two versions of the same architecture. Fugu Max aims for the best output at the lowest cost, adding more open-weight and specialized models to its pool, including NVIDIA's Nemotron family through a collaboration with NVIDIA, and costs $2 per million input tokens and $6 per million output tokens. Fugu Ultra v2 aims for peak capability at $5 and $30, rising to $10 and $45 beyond 272,000 tokens of context. Sakana says Ultra v2's training cutoff was August 28, 2026, and that Claude Fable 5, Claude Fable 5.1 and GPT-6 Astra are not in its pool. [1][4]

Claims that are hard to check

The benchmark claims are Sakana's own. It reports that Ultra v2 scored 48.3 on Chartography against 27.3 for Claude Opus 5, and 74.3 on DeepSWE, and that Max's output price is 40 to 60 percent below Sonnet 5, GPT-5.6 Terra and Kimi K3. One benchmark on which both versions lead, SWEFish, is Sakana's internal test. [1]

Sakana's product FAQ says the specific models Fugu selects, and how it coordinates them, are proprietary and not exposed by design. The service is not yet available in the EU or EEA while Sakana works toward compliance with GDPR and other EU rules. [4]

Atlas interpretation: A Fugu customer buys a routing decision it cannot inspect. That leaves Sakana's own benchmark tables as the main evidence of quality, and it means the models behind a request can change with each new Fugu version. Sakana presents that swappable pool as protection against depending on any one provider; for a buyer it is also one more layer between the request and whoever actually answered it. [1][4]