Google unveils Gemini 3, its ‘most intelligent’ AI to date

• Photo from Google

By TechWatch PH

Google has unveiled Gemini 3, calling it its most intelligent and capable AI model to date, marking what the company describes as a major step toward artificial general intelligence.

The new system powers upgrades across Google’s consumer, developer, and enterprise products and is designed to help users learn, build, and plan with deeper reasoning, stronger multimodal understanding, and more agent-like capabilities.

At the center of the rollout is Gemini 3 Pro, which Google says surpasses the performance of Gemini 2.5 Pro across major benchmarks. It tops the LM Arena leaderboard with a breakthrough 1501 Elo score and demonstrates what the company calls “PhD-level” reasoning on some of the field’s toughest evaluations.

Gemini 3 Pro delivers major gains in multimodal performance, scoring 81% on MMMU-Pro, 87.6% on Video-MMMU, and 72.1% on SimpleQA Verified, which Google says reflects meaningful improvements in factual reliability and grounded reasoning. The model is also designed to provide responses that are clear, direct, and insight-driven.

Alongside the Pro model, Google introduced Gemini 3 Deep Think, an enhanced reasoning mode built to push the system further. According to the company, Deep Think reaches 41% on Humanity’s Last Exam without tools and 93.8% on GPQA Diamond, two of the most challenging reasoning tests available today.

For coding and software development, Gemini 3 continues Google’s push toward agentic AI. The model leads the WebDev Arena leaderboard with 1487 Elo, scores 54.2% on Terminal-Bench 2.0, and reaches 76.2% on SWE-bench Verified, marking significant improvements in code reliability and problem-solving.

Google is pairing the model with Google Antigravity, a new agent-first development environment where AI agents can plan and execute full end-to-end tasks directly across the editor, terminal, and browser. Instead of simply completing prompts, the agent is designed to autonomously run workflows, test code, and self-validate outputs while the developer remains in control.

Gemini 3 is also positioned as a powerful multimodal learning companion, with capabilities that span text, images, audio, video, and code. Google highlights how the model can break down long materials into visual summaries, extract structure from complex documents, and analyze video content to generate actionable insights.

In Search, Gemini 3 now powers the new AI Mode, producing richer answers, interactive visuals, and more dynamic layouts.

Google says Gemini 3 is its most secure model yet, undergoing its most extensive set of evaluations to date. The system has been reinforced against prompt-injection attacks, tuned to reduce sycophancy, and tested with external safety groups as part of the company’s broader Frontier Safety Framework.

Gemini 3 is now rolling out across Google’s ecosystem. Users can access it through the Gemini app, while Google AI Pro and Ultra subscribers can use it inside AI Mode in Search. Developers can build with Gemini 3 through the Gemini API in AI Studio, the Gemini CLI, and Google Antigravity. Enterprise customers can deploy it via Vertex AI and Gemini Enterprise. Deep Think will be released after final safety reviews.

Google says more Gemini 3 models are on the way as it continues expanding what users can learn, build, and create with its next-generation AI.