Gemini 3.1 Pro Lands with Surprise Benchmarks and SVG Prowess, Alongside New Music AI and Security Concerns

Google has unexpectedly launched Gemini 3.1 Pro, a new iteration of its flagship AI model, without prior leaks. The model has quickly positioned itself as a top performer across most benchmarks, reportedly achieving significant leads in certain tests, including a 77.1% improvement in Arc AGI 2, an AGI-focused benchmark. While acknowledging that “benchmarks increasingly say less,” a standout feature highlighted is Gemini 3.1 Pro’s exceptional capability in generating SVG animations, demonstrating a marked improvement over its predecessor. Beyond this visual prowess, the model integrates enhanced frontend development features, exemplified by its canvas-based application generation. However, an attempt to generate a sophisticated running application using HTML, CSS, and JS yielded results described as “quite regular” and “much worse than expected,” despite incorporating design concepts like Glassmorphism and proposing unique metrics.

Concurrent with the Gemini 3.1 Pro release, Google also introduced Lydia 3, an AI model designed for generating 30-second music tracks. The limited duration has drawn some surprise, though it is speculated to be for short-form content like reels or stories. An initial test to create a rap song about programmers superior to AI resulted in an English track despite a Spanish prompt, indicating a localization challenge. Separately, Google Cloud reported detecting a “model extraction attack” targeting Gemini, where an adversary allegedly made 100,000 prompts to replicate its responses. Industry observers, however, express skepticism that 100,000 prompts would be sufficient for a full model clone, suggesting such attempts might instead focus on model distillation or improvement for competitor models.