Tokyo-based lab Sakana AI unveiled two new models on 11 September 2026: Fugu Max and Fugu Ultra v2. Rather than a single trained network, both rely on learned orchestration that dynamically routes tasks to a pool of specialised models – all served through one OpenAI-compatible endpoint.
A multi-agent system delivered as one model
Sakana describes Fugu as “a multi-agent system, delivered as one model.” Instead of depending on a single proprietary frontier model, the architecture coordinates a swappable pool of open-weight and specialised models, sending each task to the leanest model able to solve it. The company grounds the approach in two ICLR 2026 papers, TRINITY and Conductor. Fugu Ultra v2 accepts text, images and files such as PDFs and offers a context window of roughly one million tokens.
Cheaper than frontier rivals
Fugu Max targets cost efficiency at two US dollars per million input tokens and six dollars per million output tokens – which Sakana says is 40 to 60 percent below models such as Sonnet 5, GPT 5.6 Terra and Kimi K3. Fugu Ultra v2, by contrast, is tuned for peak capability.
Benchmarks are vendor claims
The headline scores come from Sakana itself and are partly internal metrics. On the visual test “Chartography,” the lab reports 48.3 points for Fugu Ultra v2 against 27.3 for Anthropic’s Opus 5 and 29.5 for Fable 5. Independent verification is still pending, so the figures should be read with caution.
Source: Sakana AI – Introducing Fugu Max and Fugu Ultra v2
