What's New in Google Gemini 3
- Oliver Xiao
- Jan 20
- 2 min read
Released in November 2025, Google Gemini 3 is a significant upgrade with "agentic" capabilities, deep reasoning, and interactive output. It replaces the previous Gemini 2.5 series with higher performance in coding, mathematics, and multimodal understanding.
Core New Features
Generative UI (Interactive Outputs): Gemini 3 can create dynamic, clickable layouts, interactive charts, and simulations tailored to a request.
Agentic Capabilities: This model can plan and execute multi-step workflows autonomously, such as triaging emails or managing complex tool-use tasks.
Deep Think Mode: Available for Ultra subscribers, this mode allows the model to "think longer" for complex problems, improving its score on benchmarks like ARC-AGI-2.
Nano Banana Pro: A new dedicated image generation and editing model integrated with Gemini 3, allowing for precise image edits.
Model Variants & Availability
Gemini 3 Pro: The primary high-intelligence model for complex reasoning and advanced coding. It holds a top position on the LMArena Leaderboard with a score of 1501.
Gemini 3 Flash: The new default model for the Gemini app and Google Search, offering "Pro-grade" reasoning at much higher speeds and lower costs.
Gemini 3 Deep Think: A specialized reasoning mode for difficult academic and scientific problems.
Developer & Professional Tools
Google Antigravity: A new agentic development platform where Gemini 3 can write code in an editor, run commands in a terminal, and browse the web to debug or validate its work.
"Vibe Coding": This allows users to build entire apps, games, or high-fidelity UI prototypes from a single natural-language prompt.
Expanded Context: Supports a 1 million token context window, allowing it to process entire codebases or long-form documents without losing coherence.
Benchmark Improvements (vs. Gemini 2.5 Pro)
Benchmark | Gemini 2.5 Pro | Gemini 3 Pro |
Humanity's Last Exam (Advanced Reasoning) | 21.6% | 37.5% |
GPQA Diamond (PhD-level Knowledge) | 88.3% | 91.9% |
SWE-bench Verified (Agentic Coding) | 59.6% | 76.2% |
MMMU-Pro (Multimodal Reasoning) | 68.0% | 81.0% |
Comments