Google announces Gemini 3, the most intelligent AI model according to the company. The model is expected to excel in reasoning ability, multimodal skills, and development.
With Gemini 3, Google introduces a new AI model that, according to the company, performs significantly better than its predecessor. The model processes not only text but also images, audio, video, and code in a single multimodal approach with a maximum token count of one million.
Within the new Gemini 3 family, Google is launching Gemini 3 Pro as a preview in various Google products. Gemini 3 promises multimodal capabilities, improved reasoning ability, and an active partner for developers. Gemini 3 Pro will appear as an option worldwide in the Gemini app. Deep Think is currently only available to testers in the United States.
Improved performance
According to Google, Gemini 3 Pro is the most powerful model in the Gemini series to date. It scores better than its predecessor Gemini 2.5 Pro on all major AI benchmarks, including LMArena, GPQA, and MathArena.
In terms of multimodal benchmarks, the model achieves 87.6 percent on Video-MMMU and 72.1 percent on SimpleQA Verified, which Google says indicates higher reliability for factual information.
The model can combine and process text, images, video, audio, and code. As a result, Google claims it can handle more complex tasks, such as analyzing scientific papers, generating visualizations, and even creating personalized training schedules based on video footage.
Deep Think
Additionally, Google is adding a new mode to Gemini 3: Deep Think. This mode focuses on solving more complex problems through improved reasoning capabilities and multimodal understanding.
According to Google, Gemini 3 Deep Think surpasses the already high performance of Gemini 3 Pro. The model scored 41 percent on Humanity’s Last Exam without aids and 93.8 percent on GPQA Diamond, a benchmark focused on advanced knowledge questions.
Furthermore, the model achieved 45.1 percent on ARC-AGI-2 (including code execution), a test that evaluates the problem-solving ability of AI models in terms of reasoning and generalization.
The feature is currently available to safety testers and will later be rolled out to users of the Google AI Ultra subscription.
For Developers
Google Antigravity is a new development platform that employs AI agents to autonomously perform complex software tasks. The platform utilizes Gemini 3 and offers developers a task-oriented development environment.
read also
Gemini 2.5 Outperforms Nearly Every Human Team in Programming Competition
Antigravity combines Gemini 3’s reasoning and programming functions with an integrated development environment. AI agents are capable of independently writing, executing, and validating code within this environment. They have direct access to the editor, terminal, and browser to handle end-to-end software processes.
In addition to Gemini 3 Pro, which forms the central reasoning model, the platform also uses Gemini 2.5 Computer Use for browser interaction and the image editing model Nano Banana (Gemini 2.5 Image).