Sakana AI Launches Fugu Max and Fugu Ultra v2 for Cheaper, Stronger Multi-Agent Orchestration

AITopTools Editorial TeamSeptember 11, 2026

What changed

Sakana AI has released Fugu Max and Fugu Ultra v2, two models built on the same system for coordinating multiple AI models. Fugu Max sends tasks to smaller, specialized open models, including NVIDIA Nemotron, while Fugu Ultra v2 is aimed at higher capability and reported scores of 48.3 on Chartography and 74.3 on DeepSWE.

What this means for you

Fugu Max is priced at $2 per million units of text processed for input and $6 for output, offering a potentially lower-cost way to handle complex tasks with several AI models. The release is available now, but the supplied information does not specify access details or explain what the reported scores mean in everyday use.

Related AI tools

Explore directory listings connected to the products, companies, and workflows in this story.

UltraBrainstomer Ultramail IMyFone UltraRepair UltraBrainstormer

Related AI news

Ideas & discoveriesSep 12, 2026

Fly Language Model (FLM) Wires the Full Fruit Fly Connectome Into a Frozen 1.2B LLM, and Its Own Controls Show the Wiring Does Not Help

This is mainly a research result rather than a ready-to-use product: the evidence suggests that adding biological wiring does not automatically improve a language model. The code is MIT-licensed and can be run locally, but the reported benefit is small and the graph-based design performed worse than its simpler control.

MarkTechPostSee why it matters

Safety & rulesSep 12, 2026

Anthropic CEO says it’s time to pump the brakes on AI

Independent evaluators may get a stronger role in checking whether Anthropic follows its safety commitments. This is a proposed approach rather than a new consumer feature, so its practical impact will depend on how the evaluations work and what Anthropic does with their findings.

The Verge AISee why it matters

New AI featuresSep 12, 2026

GPT-6 Astra appears to show a "step change" in spatial reasoning based on early benchmarks

The result could matter for future robots that need to handle objects in the real world, but it is based on an early, specialized test. The metadata does not say whether GPT-6 Astra is publicly available.

The DecoderSee why it matters