Google has announced the release of Gemini 3, the latest and most capable model in its rapidly evolving artificial intelligence family. The launch marks another milestone in the company’s effort to scale advanced AI systems across its products and services, barely two years after the introduction of the first Gemini model.

According to Google, the Gemini ecosystem has seen rapid adoption. Its AI Overviews feature now reaches about two billion users each month, the standalone Gemini app records more than 650 million monthly users, and over 13 million developers have experimented with the model family. The company adds that more than 70 percent of its cloud customers already rely on Gemini-powered tools.

Gemini’s development follows a full-stack AI strategy built on Google’s proprietary infrastructure, model training research, and product integration. Each new generation has expanded the system’s abilities. Gemini 1 introduced native multimodality and support for long-context processing, while Gemini 2 strengthened reasoning functions and laid the groundwork for agentic capabilities. Gemini 2.5 Pro later topped the LMArena benchmark for more than half a year.

Gemini 3 is positioned as the most comprehensive upgrade in the series. Google describes the model as better equipped to understand nuance, interpret user intent, and handle complex reasoning tasks with fewer prompts. The model launches across Google’s major platforms, including AI Mode in Search, the Gemini app, AI Studio, Vertex AI, and the company’s new agent-centric development platform, Google Antigravity.

One of the most significant advancements is the model’s performance on widely recognised AI benchmarks. Gemini 3 Pro achieved an Elo score of 1501 on LMArena and posted strong results on tests such as Humanity’s Last Exam and the GPQA Diamond benchmark, highlighting substantial improvements in advanced reasoning. It also delivered state-of-the-art results in multimodal evaluation frameworks, including MMMU-Pro and Video-MMMU.

The model’s “Deep Think” mode, currently in limited testing, is designed for even more demanding reasoning tasks. Early evaluations show it outperforming the main Gemini 3 Pro version on a range of frontier tests, including ARC-AGI-2, which examines a model’s ability to solve novel problems.

Beyond research benchmarks, Google says Gemini 3 is intended to support learning, planning, and creative development. With a one-million-token context window and expanded multimodal understanding, the system can analyse academic content, decipher handwritten notes, interpret lengthy videos, and generate learning tools such as visualisations or interactive flashcards. It can also analyse sports footage and propose training plans based on observed performance.

For developers, Gemini 3 brings improvements in zero-shot generation, web development, and agentic coding. It tops the WebDev Arena leaderboard and shows progress in computer-use tasks, including terminal operations and browser-based workflows. The company’s new Antigravity platform aims to elevate AI from a code-assist tool to an autonomous development agent capable of planning, executing, and validating software tasks.

Google emphasises that Gemini 3 underwent its most extensive safety evaluations to date. The model reportedly shows reduced vulnerability to prompt manipulation, improved resistance to cyber misuse, and lower tendencies toward sycophancy—an issue where AI models mirror user opinions uncritically. The company worked with external safety experts and government-linked organisations to assess its reliability.

Gemini 3 will be available globally across Google’s consumer and enterprise offerings. The enhanced Deep Think mode is expected to roll out later following additional safety testing. Google says further additions to the Gemini 3 series are planned in the coming months as it continues expanding the model’s capabilities and applications.

FAQ

What is Gemini 3?

Gemini 3 is Google’s latest artificial intelligence model, designed with enhanced reasoning, multimodal processing and agentic capabilities.

How does it differ from earlier versions?

It delivers stronger reasoning performance, better contextual understanding, improved multimodal analysis and broader product integration compared to Gemini 1, 2 and 2.5.

Where can users access Gemini 3?

The model is available through Google Search’s AI Mode, the Gemini app, AI Studio, Vertex AI, and Google’s new Antigravity developer platform.

What is Deep Think mode?

Deep Think is an advanced reasoning mode that boosts Gemini 3’s performance on complex decision-making and theoretical tasks. It is currently being tested before wider release.

Is Gemini 3 safe to use?

Google says the model underwent extensive safety evaluations, including tests by external experts, to minimise risks such as misuse, bias and prompt manipulation.