Google rolls out Gemini 3 with record reasoning scores and new antigravity coding platform

Google has introduced Gemini 3, positioning it as its most capable foundation model so far and making it available immediately on the Gemini app and through AI Mode in Google Search. The launch arrives only months after Gemini 2.5 and underscores the accelerating pace of top-tier AI model releases, following recent upgrades from OpenAI and Anthropic.

The company is also preparing a research-heavy variant called Gemini 3 Deepthink, expected to reach Google AI Ultra subscribers once additional safety checks are completed. Google highlights a major leap in reasoning performance with this generation, which is reflected in early independent benchmarks. Gemini 3 recorded a 37.4 score on the Humanity’s Last Exam test, the highest performance registered so far, and also led user-rated evaluations on LMArena.

Google notes that the Gemini ecosystem continues to scale rapidly, with more than 650 million people using the app monthly and over 13 million developers integrating Gemini tools into their work. The new model is already live across the Gemini API, AI Studio, Vertex AI and Gemini Enterprise, while the enhanced Gemini 3 Pro tier is being previewed by AI Ultra subscribers and select business customers.

A major addition to this release cycle is Google Antigravity, a coding environment built around Gemini 3. Designed for agentic development, it brings together a prompt interface, a command-line view and an in-browser window that displays real-time outcomes of code changes. The platform can plan, generate, test and verify code with minimal human intervention, offering structured progress updates directly within Chrome. Antigravity is currently available as a free public preview for Windows, macOS and Linux.

Google plans to expand access further in the coming months, beginning with upgraded AI Mode capabilities for both free and paid users in the US.