SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

Gemini 3 - Google's New Generation Multimodal Understanding and Reasoning AI Model

Author: AI小集 · Original source: AI工具集

What is Gemini 3

Gemini 3 is Google's latest AI model, hailed as the world's most advanced multimodal understanding and reasoning model. The model possesses powerful reasoning capabilities, setting multiple benchmark records, such as topping the LMArena Leaderboard with a high score of 1501 Elo. Gemini 3 supports multimodal inputs, including text, images, videos, etc., and can handle complex problems to provide reliable answers. The model introduces a deep thinking mode to further enhance its ability to solve complex problems. Gemini 3 can be used for learning and knowledge acquisition, helping developers efficiently build applications.

Users can now use Gemini 3 to build inGoogle AI Studio, Vertex AI, Gemini CLI, and Google's newly launched agent development platform Google Antigravity. The model already supports third-party platforms, such as Cursor、GitHub、JetBrains、Manus and Replit, providing developers with a wide range of choices to build and develop applications.

Gemini 3

Main Features of Gemini 3

  • Powerful Reasoning Capabilities: Gemini 3 Pro reaches a top level in reasoning capabilities, setting multiple benchmark records, such as topping the LMArena Leaderboard with 1501 Elo points, demonstrating PhD-level complex problem-solving ability.
  • Multimodal Understanding: Supports multiple modal inputs such as text, images, and videos. For example, it scored 81% and 87.6% on the MMMU-Pro and Video-MMMU tests respectively, and can parse complex charts and dynamic video streams.
  • Deep Thinking Mode: Gemini 3 Deep Think mode further enhances reasoning capabilities, demonstrating stronger complex problem-solving ability.
  • Learning and Knowledge Acquisition: Helps users learn new knowledge, such as interpreting handwritten recipes, generating interactive learning tools, supporting video content analysis, and generating training plans.
  • Development and Building: As Google's strongest programming model, it supports zero-shot generation and complex prompt processing, significantly improving development efficiency.
  • Planning and Task Management: Agent capabilities are significantly improved, enabling long-horizon planning and task management.
  • Brand-New Development Experience: Combined with the Google Antigravity platform, it enables end-to-end software development automation, supporting multi-platform development such as Google AI Studio, Vertex AI, and more.
  • Safety and Reliability: Undergone comprehensive safety evaluations, reducing sycophantic behavior, enhancing resistance to prompt injection, improving cyberattack protection, and ensuring factual accuracy.

Performance of Gemini 3

  • Outstanding Reasoning Ability: Gemini 3 Pro tops the LMArena Leaderboard with 1501 Elo points, demonstrating PhD-level reasoning ability. For example, it scored 37.5% on the "Humanity's Last Exam" and reached 91.9% on the GPQA Diamond test.
  • Leading Multimodal Understanding: Achieved high scores of 81% and 87.6% on the MMMU-Pro and Video-MMMU tests respectively.
  • Breakthrough in Deep Thinking Mode: Gemini 3 Deep Think mode scored 41.0% on "Humanity's Last Exam", reached 93.8% on the GPQA Diamond test, and achieved 45.1% on the ARC-AGI-2 test, significantly improving complex problem-solving ability.
  • Outstanding Mathematical Ability: Achieved a new state-of-the-art score of 23.4% on the MathArena Apex test, setting a new standard for frontier models in mathematical reasoning.
  • Improved Factual Accuracy: Achieved a score of 72.1% on the SimpleQA Verified test, demonstrating significant progress in providing accurate information.
  • Significantly improved development efficiency: Topped the WebDev Arena leaderboard with 1487 Elo points, significantly boosting developer efficiency and supporting complex web UI and application development.
  • Enhanced tool use capabilities: Scored 54.2% on the Terminal-Bench 2.0 test and greatly surpassed its predecessor on the SWE-bench Verified test, delivering outstanding performance.
  • Improved long-horizon planning capabilities: Topped the Vending-Bench 2 test, demonstrating exceptional long-horizon task planning and decision-making coherence.
Gemini 3

How to use Gemini 3

  • General users: Through Gemini directly, or experience it in AI Mode in Search with Google AI Pro and Ultra subscription services.
  • Developers: Through Google AI Studio、Vertex AI、Gemini CLI or Google's brand-new agent development platform Google Antigravity.
  • Enterprise users: Access through the Vertex AI platform or Gemini Enterprise.
  • Deep Think mode: In the coming weeks, Google AI Ultra subscribers will be able to use Gemini 3's Deep Think mode, which is currently undergoing safety evaluation.

Gemini 3 product pricing

Gemini 3.0 Pro introduces tiered pricing based on context length, as follows:

  • Tasks under 200k tokens：
    • Input price: $2.00 per million tokens.
    • Output price: $12.00 per million tokens.
  • Tasks over 200k tokens：
    • Input price: $4.00 per million tokens.
    • Output price: $18.00 per million tokens.

Gemini 3 application scenarios

  • Learning and education: The model can integrate multimodal information and generate interactive learning tools to help users efficiently learn new knowledge.
  • Development and programming: As a powerful programming model, it supports zero-shot generation and complex prompt handling, significantly improving development efficiency.
  • Task planning and management: Gemini 3's Agent capabilities support long-horizon task planning, helping users manage complex tasks and daily affairs.
  • Content creation: Gemini 3 can generate high-quality creative content, such as poems, stories, and game code, supporting creative expression.
  • Knowledge management and search: Provides intelligent generative UI in Google Search, helping users obtain and integrate information more efficiently.