Gemini 3 Release Notes

Google DeepMind has introduced Gemini 3, its most intelligent model to date, designed to integrate multimodal understanding, advanced reasoning, and agentic capabilities to help users learn, build, and plan. The release includes Gemini 3 Pro, a high-performance model available across Google products, and Gemini 3 Deep Think, an enhanced reasoning mode for highly complex tasks.

State-of-the-Art Reasoning and Benchmarks

Gemini 3 Pro significantly outperforms Gemini 2.5 Pro across all major AI benchmarks, establishing new standards in reasoning, mathematics, and factual accuracy.

Performance Metrics

  • General Intelligence: Gemini 3 Pro tops the LMArena Leaderboard with a score of 1501 Elo.
  • PhD-Level Reasoning: The model achieves 37.5% on Humanity’s Last Exam (without tools) and 91.9% on GPQA Diamond.
  • Mathematics: It sets a new frontier model state-of-the-art with 23.4% on MathArena Apex.
  • Multimodal Reasoning: Gemini 3 Pro scores 81% on MMMU-Pro and 87.6% on Video-MMMU.
  • Factual Accuracy: It achieves 72.1% on SimpleQA Verified.

Gemini 3 Deep Think

Gemini 3 Deep Think is a specialized mode that further pushes reasoning boundaries. In testing, it achieves:

  • Humanity’s Last Exam: 41.0% (without tools).
  • GPQA Diamond: 93.8%.
  • ARC-AGI-2: 45.1% (with code execution, ARC Prize Verified), demonstrating a capacity for solving novel challenges.

Multimodal Learning and Search Integration

Gemini 3 combines a 1 million-token context window with advanced vision and spatial understanding to synthesize information across text, images, video, audio, and code.

Educational Applications

Gemini 3 can translate handwritten recipes in multiple languages into cookbooks or analyze academic papers and long video lectures to generate interactive flashcards and visualizations. It can also analyze sports videos (e.g., pickleball) to provide form improvements and training plans.

AI Mode in Search

For the first time, Gemini is shipping in Google Search on day one. AI Mode in Search utilizes Gemini 3 to create generative UI experiences, including immersive visual layouts and interactive simulations generated on the fly based on user queries.

Agentic Coding and Google Antigravity

Gemini 3 is positioned as a powerful "vibe coding" and agentic model, designed to automate complex software tasks and increase developer productivity.

Coding Benchmarks

  • Web Development: Gemini 3 tops the WebDev Arena leaderboard with 1487 Elo.
  • Tool Use: It scores 54.2% on Terminal-Bench 2.0, testing the ability to operate a computer via terminal.
  • Coding Agents: It significantly outperforms Gemini 2.5 Pro on SWE-bench Verified, scoring 76.2%.

Google Antigravity

Google has launched Google Antigravity, an agent-first development platform. Unlike traditional AI IDEs, Antigravity's agents have direct access to the editor, terminal, and browser, allowing them to autonomously plan, execute, and validate end-to-end software tasks. This platform integrates Gemini 3 Pro, the Gemini 2.5 Computer Use model for browser control, and the Nano Banana (Gemini 2.5 Image) image editing model.

Long-Horizon Planning and Personal Agency

Gemini 3 Pro improves the ability to reliably plan over longer horizons and maintain consistent decision-making.

  • Vending-Bench 2: Gemini 3 Pro tops this leaderboard, maintaining consistent tool usage and decision-making for a simulated year of business operation without drifting off task.
  • Everyday Utility: These improvements enable the model to handle multi-step workflows, such as organizing an inbox or booking local services, via the Gemini Agent available to Google AI Ultra subscribers.

Safety and Responsibility

Gemini 3 has undergone the most comprehensive safety evaluations of any Google AI model to date. Key improvements include:

  • Reduced Sycophancy: The model is less likely to simply agree with the user.
  • Security: Increased resistance to prompt injections and improved protection against cyberattacks.
  • External Validation: Evaluations were conducted in partnership with subject matter experts, the UK AISI, and independent firms including Apollo, Vaultis, and Dreadnode.

Sources