Google DeepMindLanguage model

Gemini 3 Deep Think

Google DeepMind language model.

ARC-AGI-2

#14

of 54
10.4 pts behind #1

races led

weeks at #1

84.6

score

The 30-second version

Google DeepMind language model.

Key takeaways

Latest real results · click any race to watch it

3 Deep ThinkBest on each axis

Hard Reasoning

#1 frontier3 Deep ThinkGrok 4.3M3

The distilled version first; open any release for the deeper breakdown.

60-second briefing

3 Deep Think, explained

Animated summary · captions from the auto-written script

From Google's own guide, distilled

Structure prompts as persona, task, context and format; the format line is the most often forgotten.
Few-shot examples set the pattern; keep their formatting identical so the model copies structure, not noise.
Gemini reads images, video and PDFs natively, so attach the source instead of describing it.
Place media before the text instruction that refers to it.
Break complex jobs into a chain of smaller prompts and pass outputs forward.

About Google DeepMind

Google's AI lab behind Gemini, Veo, Imagen and Nano Banana.

Optimising for

Native multimodality (image, video, audio), very long context, science, and tight integration with Search, Workspace and Android.

London, UK / Mountain View, US · founded 2010

Closest rivals

3 Deep Think is 14th in arc-agi-2; leader is GPT-6 Astra.