
What happened
Sakana AI announced Fugu Ultra v2 and Fugu Max, saying Fugu Ultra v2 topped models like GPT-6 Astra on benchmarks such as DeepSWE, while not using those models as partners.
Why it matters
Fugu Max, which newly adds Nvidia's Nemotron open-weight series, is said to cut output prices 40–60% versus GPT‑5.6 Terra, Claude Sonnet 5 and Kimi K3.
What to watch
The claims hinge on benchmark scores, and readers should watch the API prices per 1 million tokens: Fugu Ultra v2 at $5 input and $30 output, Fugu Max at $2 input and $6 output.
WHO IT HITSEnterprise buyers and developers choosing an AI API face a new option priced and positioned against OpenAI and Anthropic models, though the accuracy of the claimed benchmark wins remains to be seen.
Ask the AI about this article →
Summaries like this, in your inbox every morning.
Sakana Fugu is an AI system that Sakana AI began offering generally in June. It selects models based on the difficulty of a user's request and, when needed, builds and coordinates a team of dedicated AI agents. Because the processing stays internal, users can treat it like a single AI model through the API.
The latest version, Fugu Ultra v2, sits at the top of the Fugu lineup. On some benchmarks, it is said to outscore GPT-6 Astra, Claude Fable 5.1 and Claude Fable 5, even though those three models are not among its partners. This suggests the gains come from how Fugu combines and orchestrates other models rather than from any single partner model.
Fugu Max, by contrast, focuses on cost performance, newly adding Nvidia's Nemotron open-weight series and others as partners. How much these claimed advantages translate into real-world use will likely hinge on whether the benchmark wins hold up in practice and whether the lower output prices attract developers away from rival models.
For example, today's edition would include:
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. Q&As are published on this page for other readers too.
Much of the attention on AI infrastructure buildouts is now tied to sheer compute power, with dominance define…

Barron's reported September 10 that Kepler Computing emerged from stealth with a memory architecture using fer…

Dynatrace acquired Arize AI, adding AI observability, evaluation and agent monitoring to its application obser…
Reuters reported September 10 that inference-chip startup d-Matrix will use Nvidia's NVLink Fusion to connect…

A Daily Dose of Data Science test kept LoRA adapters separate from a shared 7B base model, cutting 100 fine-tu…

A report by Spencer Kitts, Thomas Larsen and Sydney Von Arx says an OpenAI agent swarm very likely ran an atta…
