Google launches Gemini 3 AI with multimodal & reasoning boost

Google has unveiled Gemini 3, its latest artificial intelligence model, which is now available to users, developers and enterprises across multiple products and platforms. The release marks a significant expansion in the company's AI portfolio, introducing advanced reasoning and multimodal understanding for a wide range of tasks from learning and research, to complex coding and software development.

Model capabilities

Gemini 3 is described as Google's most capable model to date, supporting inputs and outputs across various media such as text, images, video, audio and code. The model demonstrates advanced performance in reasoning, planning, and context understanding, enabling it to tackle complex challenges in both professional and consumer domains.

Benchmark results show Gemini 3 Pro outperforming its predecessor, Gemini 2.5 Pro, across key evaluation metrics. It achieved a leading score of 1501 Elo on the LMArena Leaderboard and delivered top marks on exams assessing reasoning, such as a 37.5% result without tools on Humanity's Last Exam and 91.9% on GPQA Diamond. Additionally, it set a new record of 23.4% on MathArena Apex for mathematical problem-solving.

Multimodal strengths

The model also excels in multimodal benchmarks, registering 81% on MMMU-Pro and 87.6% on Video-MMMU, pointing to improved performance in tasks that require cross-referencing information from diverse sources. On the SimpleQA Verified benchmark, Gemini 3 scored 72.1%, highlighting improvements in factual accuracy and reliability.

Gemini 3's expanded context window-capable of handling up to one million tokens-enables it to synthesise and analyse large volumes of data for more comprehensive outputs, including summarising lengthy research papers or generating interactive educational materials.

New features

The launch is accompanied by the introduction of the Gemini app's redesigned interface and new features such as dynamic visual layouts. An experimental feature, Gemini Agent, has been introduced for handling tasks that require multiple steps, such as managing emails or booking services.

Gemini 3 Deep Think, an enhanced reasoning mode, pushes the capabilities of Gemini 3 even further, offering improved performance on complex tasks with more nuanced understanding. In initial testing, this mode registered a 41.0% score without tools on Humanity's Last Exam and 93.8% on GPQA Diamond, along with a 45.1% result on the ARC-AGI-2 coding benchmark.

Developer tools

Google introduced a new agentic development platform called Google Antigravity, designed to help developers create and manage software at a higher level of abstraction. This platform combines Gemini 3's reasoning and coding abilities, enabling agents to autonomously plan and execute software development tasks, validate code, and operate integrated development environments, terminals, and browsers. Gemini 3 Pro and other AI models are available through Antigravity, as well as conventional interfaces such as AI Studio and Gemini CLI.

The company has highlighted integration with third-party developer platforms, including Cursor, GitHub, JetBrains, Manus, and Replit, broadening the reach of Gemini 3's capabilities in the software development ecosystem.

Enterprise offering

For enterprises, Gemini 3 is accessible through Google's Vertex AI and Gemini Enterprise solutions. The model's extended planning abilities have been demonstrated through its leading performance on Vending-Bench 2, simulating business management tasks over prolonged periods and maintaining consistent decision-making and tool use.

Gemini 3's agentic features allow it to carry out extended workflows and optimise processes such as booking appointments, managing workflows, or analysing business operations end-to-end, all while remaining under user control.

Safety and evaluations

Google states that Gemini 3 is its most secure model to date, following extensive safety evaluations involving in-house experts and external organisations. Its safeguards include reduced susceptibility to prompt injection attacks, resistance to sycophantic responses, and improved protection against misuse.

Safety reviews have included collaboration with the UK's Artificial Intelligence Safety Institute and independent assessments by industry specialists.

"It's amazing to think that in just two years, AI has evolved from simply reading text and images to reading the room," said Sundar Pichai, CEO, Google and Alphabet.