Subscribe
Share
Share
Embed
Sakana Fugu Ultra 1.1: Multi‑Agent “Council of Models” That Beats Fable 5 (Goldy Bench Tests)
The video reviews Sakana Fugu Ultra 1.1, a new Japan-based model that benchmarks above Fable 5 and uses a “council of models” approach: one prompt is routed by an orchestrator to up to three expert models whose outputs are merged into a single build. The creator tests it on Goldy Bench by generating multiple one-shot game builds (e.g., open-world and flight simulator) to evaluate planning, logic, coding, and UI, noting the main drawback is slow generation (about 15–25 minutes) with no back-and-forth. Side-by-side comparisons show Fugu Ultra producing smoother, less buggy, better-looking results than Fable 5. The script also mentions regional blocking in the EU/UK, alternatives like OpenRouter Fusion and Hermes mixture-of-experts, and integrating everything into an Agent OS with saved workspace, plus access options via console.sakana.ai or OpenRouter and a pitch for the AI Profit Boardroom.
The video reviews Sakana Fugu Ultra 1.1, a new Japan-based model that benchmarks above Fable 5 and uses a “council of models” approach: one prompt is routed by an orchestrator to up to three expert models whose outputs are merged into a single build. The creator tests it on Goldy Bench by generating multiple one-shot game builds (e.g., open-world and flight simulator) to evaluate planning, logic, coding, and UI, noting the main drawback is slow generation (about 15–25 minutes) with no back-and-forth. Side-by-side comparisons show Fugu Ultra producing smoother, less buggy, better-looking results than Fable 5. The script also mentions regional blocking in the EU/UK, alternatives like OpenRouter Fusion and Hermes mixture-of-experts, and integrating everything into an Agent OS with saved workspace, plus access options via console.sakana.ai or OpenRouter and a pitch for the AI Profit Boardroom.
Creators and Guests
What is AI News Today | Julian Goldie Podcast?
Latest Podcast