Google launches Gemini 3 model with advanced reasoning and global leadership in AI benchmarks

Google announced this Tuesday (18) the arrival of the Gemini 3, described by the company’s leadership as its most sophisticated artificial intelligence model to date, designed to understand nuances and perform tasks with unprecedented autonomy. The new generation of technology promises to transform digital interaction by integrating superior multimodal capabilities and in-depth logical reasoning directly into the company’s products, from the search engine to complex software development tools. The update marks a new phase in the fight for dominance in generative AI, focusing not just on responding to commands, but on acting as a thought partner.

Main innovations presented at launch:

• Aprofundado Reasoning: Introdução of Deep Think mode for solving complex problems.
• Agentes autonomy: Capacidade improved planning and execution of multi-step tasks.
• Google Antigravity: Nova developer-focused platform that integrates coding agents.
• Recorde performance: Liderança confirmed in the LMArena ranking and in academic tests.

Advances in logical reasoning and multimodal capabilities

The big difference of Gemini 3 lies in its “state of the art reasoning” architecture, which allows the model to process information with a greater level of depth than previous versions. Segundo Sundar Pichai, CEO of Google and Alphabet, technology has evolved from simply reading text and images to understanding the context and intent behind user requests, eliminating the need for overly detailed prompts.

The company also unveiled “Gemini 3 Deep Think”, an enhanced reasoning mode that achieved impressive scores in high difficulty assessments, demonstrating the ability to deal with challenges that require sequential logic and interpretation of abstract data.

Performance tests indicate a significant qualitative leap compared to the 2.5 Pro model:

• Score of 1501 Elo in LMArena’s Leaderboard.
• 91.9% approval on the GPQA Diamond benchmark.
• 87.6% accuracy on the Video-MMMU for multimodal understanding.

New Google Antigravity platform for developers

Simultaneously with the launch of the model, the Google Antigravity platform was presented, described as an “agent-first” development environment. The tool allows programmers to operate at a task-oriented level, where AI goes from being just a code-completion assistant to an active partner that plans and executes end-to-end workflows.

With direct access to the editor, terminal and browser, Gemini 3 agents within Antigravity can develop, test and validate end-to-end applications. The integration aims to increase productivity by delegating repetitive and complex processes to artificial intelligence, maintaining human supervision for strategic and creative decisions in software development.

Direct integration into search and coding capabilities

The update also directly impacts the end consumer with the immediate implementation of Gemini 3 in the “AI Mode” of Busca of Google.

For the first time, a model of this magnitude is made available in the search engine on launch day, allowing the generation of dynamic interfaces and immersive visual layouts in real time based on queries.

Additionally, the “vibe coding” capability has been enhanced, allowing the model to generate rich, interactive web visualizations based on natural language instructions.

Superior performance on math and science tests

Factual accuracy and mathematical ability were points of focus in the development of this generation.

The model achieved a new high of 23.4% in MathArena Apex, setting a record for frontier models in advanced mathematical resolution.

In fact-checking tests such as the SimpleQA Verified, the Gemini 3 achieved 72.1%, suggesting a reduction in “hallucinations” common in generative models.

This data reinforces the usefulness of the tool for academics and scientists who require reliability in synthesizing large volumes of technical data.

Long-term planning ability

One of the more technical developments detailed by the Google DeepMind team is model consistency across long-horizon tasks. In business simulations such as the Vending-Bench 2 benchmark, the Gemini 3 was able to manage virtual operations for extended periods without deviating from the initial objective, something that previous models had difficulty maintaining. Isso suggests that AI can be used with greater security to organize personal workflows, such as scheduling services and managing emails, maintaining the coherence of actions from the beginning to the end of the process under the user’s supervision.

Availability and immediate access to resources

Gemini 3 Pro is now available to developers through Google AI Studio and Vertex AI, as well as third-party platform integrations.

For the general public, the features begin to be distributed today in the Gemini and Busca applications, with advanced features reserved for AI Ultra plan subscribers.