Google Announces Gemini 3: Next-Gen AI Model Leads Industry in Reasoning

On November 18, 2025, Google announced its next-generation AI model, Gemini 3. The company positions it as its “most intelligent model,” achieving industry-leading results in reasoning capabilities, multimodal understanding, and coding performance¹.

Gemini 3 Pro achieved a historic 1501 Elo score on the LMArena leaderboard—the first model to surpass 1500. It scored 91.9% on GPQA Diamond, which measures PhD-level scientific reasoning, and 37.5% on Humanity’s Last Exam without tool usage¹. The Gemini app has reached 650 million monthly users, while AI Overviews serves 2 billion monthly users, demonstrating the widespread adoption of Google’s AI products¹.

Improved Reasoning Performance

Gemini 3 shows improved performance across all major benchmarks compared to its predecessor, Gemini 2.5 Pro. Key highlights include¹²:

LMArena: 1501 Elo (surpassing GPT-5.1 and Claude 4.5 Sonnet)
GPQA Diamond: 91.9% (PhD-level scientific reasoning)
Humanity’s Last Exam: 37.5% (without tool usage)
MathArena Apex: 23.4% (new state-of-the-art in mathematical reasoning)
AIME 2025: 95% (high school mathematics)
SimpleQA Verified: 72.1% (factual accuracy)

In multimodal performance, the model achieved 81% on MMMU-Pro and 87.6% on Video-MMU¹.

Deep Think Mode

Gemini 3 features an extended reasoning mode called “Deep Think.” This mode performs advanced reasoning by exploring multiple hypotheses in parallel, enabling it to handle more complex problems¹.

Deep Think mode performance¹:

Humanity’s Last Exam: 41.0% (up from 37.5% in standard mode)
GPQA Diamond: 93.8% (up from 91.9% in standard mode)
ARC-AGI-2: 45.1% (with code execution, ARC Prize verified)

Deep Think mode is currently available to Google AI Ultra subscribers following safety evaluations³.

Coding and Agent Capabilities

Gemini 3 has received high praise for “vibe coding” (prompt-based code generation). It achieved 1487 Elo on the WebDev Arena leaderboard and 76.2% on SWE-bench Verified¹.

Google Antigravity

Announced alongside Gemini 3, “Google Antigravity” is an agent-oriented development platform. It leverages Gemini 3’s advanced reasoning and tool-use capabilities to provide an environment where developers can give instructions at a higher level of abstraction¹².

Agents have direct access to the editor, terminal, and browser, autonomously planning and executing complex software tasks. The platform also integrates the Gemini 2.5 Computer Use model (for browser control) and Nano Banana (for image editing)¹.

Gemini Agent

“Gemini Agent” is available for Google AI Ultra subscribers. This agent functionality can execute multi-step tasks such as organizing Gmail and managing Google Calendar¹².

On Vending-Bench 2, which measures long-horizon planning capabilities, Gemini 3 Pro topped the leaderboard, maintaining consistent tool usage and decision-making throughout a simulated year of operation¹.

Multimodal Learning Support

Gemini 3 has the ability to understand across text, images, video, audio, and code. Utilizing its 1 million token context window, it enables learning support such as¹:

Deciphering and translating handwritten recipes
Generating interactive flashcards from academic papers and long video lectures
Sports video analysis and training plan creation

In Google Search’s AI Mode, “generative UI” powered by Gemini 3 has been introduced, dynamically generating interactive layouts and simulations based on queries¹².

Safety Initiatives

According to Google, Gemini 3 has undergone the most comprehensive safety evaluation of any Google AI model to date. The following improvements have been reported¹:

Reduced sycophancy
Improved resistance to prompt injection
Enhanced protection against misuse through cyberattacks

Early access has been provided to organizations such as UK AISI, and independent assessments have been conducted by industry experts including Apollo, Vaultis, and Dreadnode¹.

Availability

Gemini 3 is available in the following environments¹:

General users: Gemini app, Google Search AI Mode (for Pro/Ultra subscribers)
Developers: Google AI Studio, Vertex AI, Gemini CLI, Google Antigravity
Third-party platforms: Cursor, GitHub, JetBrains, Manus, Replit, and more

The model supports up to 1 million input tokens and 64K output tokens⁴.

For more details on Gemini 3, visit the official Google DeepMind blog or try it out in Google AI Studio.