Blog/News
ARC-AGI-2: Thinking mode reaches ~35%
2025-10-20On the ultra-difficult ARC-AGI-2 general intelligence test, Gemini 3.0 with thinking mode reaches ~35% accuracy, while others stay below 20%.
HLE benchmark: 32.4% top score
2025-10-20On the notoriously hard “Human Last Exam (HLE)” benchmark, Gemini 3.0 scores 32.4% — outperforming GPT-5 (high) and Grok 4.
SVG pelican output amazes the community
2025-10-20Gemini 3.0 handles images including SVGs with ease. The famous cycling pelican SVG test wowed the community with Gemini 3.0 Pro output.
Gundam & Switch controller rendering upgrade
2025-10-20Previously hard Gundam robot and Switch controller renders now look visibly upgraded — getting very close to real product photos.
Leading the coding arena
2025-10-20Gemini 3.0 Pro leads the coding arena by a wide margin.
Series leap toward super intelligence
2025-10-20Gemini 3 has become truly super‑intelligent; the series makes a leap forward.
Prompted image: Power Rangers example
2025-10-20Gemini 3.0 Pro generated an image from the prompt “the Power Rangers standing in the scene with typical poses the power rangers do”.