Key Highlights –
- Google has officially released Gemini 3 Pro, its “most intelligent” model to date, which is now available in preview across the Gemini app, Google Search, and for developers via API.
- The model features a new “Deep Think” mode that achieves state-of-the-art scores on benchmarks like Humanity’s Last Exam (41.0%) and GPQA Diamond (93.8%), pushing the boundaries of complex reasoning.
- Gemini 3 offers native multimodal capabilities, achieving high scores on video and image benchmarks (87.6% on Video-MMMU), and introduces a new agentic development platform called Google Antigravity.
Google’s highly anticipated Gemini 3 has officially launched. The latest iteration of its flagship artificial intelligence model, marking what CEO Sundar Pichai calls “a new era of intelligence.” The release, announced today by Google DeepMind CEO Demis Hassabis and Chief AI Architect Koray Kavukcuoglu, introduces Gemini 3 Pro in preview, alongside a new enhanced reasoning capability dubbed “Deep Think.”
Google claims that the new Gemini 3 is the “strongest model” for multimodality in the world, directly targeting at its competitors given how there’s an ever increasing demand for AI that can handle complex, multi-step reasoning tasks. The company states the model is now available across its product suite, including the Gemini app, Google Search, and enterprise platforms like Vertex AI.
Introducing Gemini 3 — our most intelligent model that helps you bring any idea to life.
— Google (@Google) November 18, 2025
Gemini 3 is our next step on the path toward AGI and has:
🧠 State-of-the-art reasoning
🖼️ Deep multimodal understanding
💻 Powerful vibe coding so you can go from prompt to app in one shot… pic.twitter.com/zG8r95pGcS
Deep Think and Benchmark Dominance
The headline feature of the release is Gemini 3 Deep Think, a specialized mode designed for complex problem-solving. According to Google, this mode pushes performance beyond the standard Pro model, achieving a score of 41.0% on “Humanity’s Last Exam” (without tools) and 93.8% on the GPQA Diamond benchmark. It also reportedly achieves an unprecedented 45.1% on the ARC-AGI-2 benchmark with code execution, signaling strong capabilities in solving novel, unseen challenges.
Even the standard Gemini 3 Pro model claims to outperform its predecessor, Gemini 2.5 Pro, on “every major AI benchmark.” Google reports a breakthrough Elo score of 1501 on the LMArena Leaderboard and PhD-level reasoning capabilities. In mathematics, the model set a new state-of-the-art record of 23.4% on MathArena Apex. Beyond raw numbers, Google emphasizes the model’s “nuance,” stating that Gemini 3 trades cliché and flattery for “concise and direct” responses. This shift aims to position the AI as a more effective “thought partner” for users, capable of tasks ranging from coding high-fidelity visualizations of scientific concepts to creative brainstorming.
Multimodal Mastery and Developer Tools
Gemini 3 is built to be natively multimodal, excelling at processing text, images, and video simultaneously. The model scored 87.6% on the Video-MMMU benchmark and 81% on MMMU-Pro. Google also highlighted the model’s ability to generate code for complex visualizations, such as simulating plasma flow in a tokamak, demonstrating a fusion of scientific knowledge and coding proficiency.
For developers, the launch includes immediate access to Gemini 3 via the Gemini API in AI Studio. Google also introduced Google Antigravity, a new platform specifically designed for building agentic AI applications, and a new Gemini CLI (Command Line Interface). Enterprise customers can access the model starting today through Vertex AI.
Google notes that while Gemini 3 Pro is rolling out immediately, the advanced “Deep Think” mode will initially be available to safety testers before expanding to Google AI Ultra subscribers in the coming weeks. The company claims this is its “most secure model yet,” citing reduced sycophancy and improved resistance to prompt injection attacks
