Google has presented Gemini 3Its new generation of artificial intelligence model is designed to combine advanced reasoning, context understanding, and multimodal capabilities. The company positions it as its most capable system to date, designed to interpret nuances and solve complex tasks with less back-and-forth.
In the initial deployment, the Gemini app and the developer tools already integrate this version, while the AI Mode in Search It launches first in the United States. In Spain and the rest of Europe, Google is prioritizing access through the app, API, and cloud services, along with an expansion of languages to include Catalan, Basque, and Galician.
What does Gemini 3 bring and why is it relevant?
Gemini 3 combines in a single system the accumulated experience of previous generations and a leap in reasoning and understanding of the contextThe company maintains that the model “captures subtle clues” in creative ideas and breaks down complex problems into layers, with more accurate answers and less need to rephrase the query.
This approach results in more concise answers that prioritize the Useful information versus clichésThis is a behavior that Google had been testing to reduce flattery and improve factual accuracy. Furthermore, the multimodal nature of the model allows working with text, images, audio, and video in the same interaction.
Performance in independent and comparative tests
According to metrics published by Google and public rankings, Gemini 3 Pro He leads LMArena with 1.501 ELO points, surpassing his predecessor. In academic assessment, he achieves 37,5% without tools in Humanity's Last Exam and 91,9% in GPQA Diamond; in mathematics, he scores 23,4% in MathArena Apex.
The reinforced mode of reasoning DeepThink It takes these results even further, with 41,0% in Humanity's Last Exam and 93,8% in GPQA Diamond. ARC-AGI-2, with code execution, registers 45,1%, an indicator of progress in abstract reasoning and visual comprehension.
Multimodality and a huge context window in Gemini 3
One of the big changes is the context window of 1 million tokensThis facilitates the analysis of extensive code repositories, papers, or long videos. The model can, for example, translate and unify handwritten recipes to create a family cookbook, or convert complex materials into visualizations and interactive cards.
Multimodal reading is reflected in tests such as MMMU-Pro (81%) and Video-MMMU (87,6%), and in SimpleQA Verified (72,1%), where the focus is on greater factual accuracy. In AI Mode, integrated into Search, Google teaches interactive elements—such as widgets and simulations—to explain concepts on the fly.
Programming and agents: the impetus for Antigravity
In development tasks, Gemini 3 leads WebDev Arena with 1.487 ELOIt achieves a 54,2% in Terminal-Bench 2.0 and a 76,2% in SWE-bench Verified, which translates into greater competence in generating code, using tools and operating a terminal environment.
The most notable new feature for building agents is Google AntigravityThis platform gives agents access to the editor, terminal, and browser. With this, the system can plan and execute end-to-end software workflows, validating its own code and leaving the final review and audit to the developer.
Safety, external testing, and responsible deployment
Google claims this is its Safest model to dateIt is reinforced against flattery and more resistant to prompt injections and misuse (e.g., cyberattacks). The assessment is part of its Frontier Safety Framework and has been reviewed by external experts, including the AISI of the United Kingdom and firms like Apollo, Vaultis and Dreadnode.
Mode DeepThink It undergoes additional validation before reaching Google AI Ultra subscribers, due to the scope of its planning capabilities and tool usage. The company emphasizes a gradual and measured rollout across products and regions.
Gemini 3 Availability
El Gemini 3 Pro preview It is accessible in the Gemini app and via the developer API through AI Studio and Vertex AI. The interface adds 30 new languages, including Catalan, Basque and Galician, which facilitates its adoption in local contexts.
El AI Mode in Search It starts in the United States and will gradually expand; meanwhile, companies and technical teams can integrate Gemini 3 into their workflows with Vertex AI and Gemini Enterprise, in addition to the new CLI and the Antigravity platform for agents.
Use cases: from learning better to automating tasks
Beyond chat, the model can transform complex content into interactive experiences with graphs, calculators, or simulations. It is useful for studying technical concepts, analyzing sports matches and proposing training plans, or breaking down papers and generating navigable guides.
In their day-to-day work, agents can classify Gmail email, plan trips or execute multi-step tasks by combining reasoning, tool calls and browser control, always with human supervision to maintain control.
Google maintains that Gemini's impact is growing: "AI-powered View" arrives 2.000 million monthly usersGemini's app exceeds 650 million monthly downloads, 70% of cloud customers use its AI capabilities, and 13 million developers have built solutions with its models. These figures underpin the deployment of Gemini 3 within the product ecosystem.
With its focus on reasoning, multimodality, and agents, Gemini 3 It is emerging as the common basis of Google's AI strategy: a more direct, less accommodating platform with more context, which in Spain you can start using from the Gemini app and, if you develop, from AI Studio and Vertex AI; AI Mode Search will be activated by market.
