Sakana Fugu Ultra v2: What Changed in September 2026

In September 2026, Sakana AI followed its June Fugu launch with Fugu Max and Fugu Ultra v2. Here is what changed and what to watch.

What is new

• Fugu Max: positioned for high-volume tasks where peak reasoning is not required — the cost-efficient workhorse tier.
• Fugu Ultra v2: the heavier reasoning tier, with reported scores on benchmarks like Chartography and DeepSWE now awaiting independent replication.
• NVIDIA partnership (announced August 2026): gives Fugu Max access to Nemotron open models that would otherwise be impractical to integrate at orchestration scale.
• Technical report: Sakana published a methodology paper (arXiv) providing a baseline for third-party replication — the right move for credibility.

The competitive picture

Fugu arrived in the same season as other multi-agent plays — notably multi-agent coding environments — but targets a different layer: API-level orchestration rather than IDE-integrated workflows. Enterprise buyers evaluating long-running agentic workloads will likely compare both.

What to watch next

Three things: independent labs reproducing the benchmark scores, pricing for the above-272K-token tier (still unspecified — and long-context agentic work regularly exceeds that), and real enterprise adoption through partnerships like the SCSK/Sumitomo alliance in Japan. Read our plain-English Fugu explainer for the fundamentals, or the overview.