Google has just launched the Gemini 3 model, which focuses on improving coding, email processing, and document analysis. This model also boasts better reasoning abilities, context understanding, and the ability to incorporate graphics (diagrams, illustrations, etc.) when answering certain questions.
Google's announcement highlighted that Gemini 3 will appear in the Gemini app, the AI mode on Google Search, and a new programming platform called Google Antigravity.
"The smartest from Google"
Since its launch nearly two years ago, Google has continuously expanded Gemini to reach a wider audience. The company highlights that its AI Overviews feature currently has 2 billion monthly users, the Gemini app has 650 million monthly users, over 70% of cloud customers use AI, and 13 million developers are creating products based on Google's AI generation model.
For the first time, Google allowed users to access the Gemini 3 Pro within the Gemini app from day one, claiming it was "the company's smartest model yet." The company also rolled out the Gemini 3 Pro through AI Mode for paying users.
Tulsee Doshi, Senior Director and Head of Product at Google DeepMind, said Gemini 3 will bring the company closer to its goal of making information “widely accessible and useful,” further enhancing Google’s search engine.
Google emphasizes the Gemini 3 Pro's multimodal processing capabilities, centered around processing text, images, and audio simultaneously. For example, the model can read photos of recipes and convert them into books, or create interactive flashcards based on video lectures.
 |
The benchmark results for the Gemini 3 Pro were compiled by Google on LMArena. Image: Google . |
The announcement from Google also highlighted the Gemini 3 Pro's performance across benchmarks. For example, the model scored 1,501 points, topping the LMArena rankings. The Gemini 3 Pro's reasoning ability reached a doctoral level with a 37.5% score on the Humanity's Last Exam and 91.9% on the GPQA Diamond exam.
On the MathArena Apex scale, Gemini 3 Pro achieved 23.4%, currently the highest score among large model languages (LLMs).
Regarding multimodal reasoning capabilities, Google states that the Gemini 3 Pro achieved 81% on the MMMU-Pro scale and 87.6% on the Video-MMMU scale. The model also scored 72.1% on the SimpleQA Verified accuracy testing platform.
"This demonstrates that the Gemini 3 Pro is capable of solving complex problems across a wide range of fields, from science to mathematics, with high reliability," Google emphasized.
With the Gemini 3 Deep Think version, scores on the Humanity's Last Exam and GPQA Diamond tests were even higher than Gemini 3 Pro, at 41% and 93.8% respectively. The model also achieved 45.1% on ARC-AGI, a general AI metric.
What can Gemini 3 do?
Improvements in Gemini 3 will appear across many Google products. The Dynamic View feature, currently being tested in Gemini, creates an interactive experience for certain queries, such as magazines with illustrations, brief information, or customized interfaces on demand.
"Gemini 3 can even analyze your pickleball game video, identify areas for improvement, and create a training plan to enhance your overall technique," Google emphasized.
 |
The Gemini 3 Pro's advanced encoding capabilities help produce better images. Photo: Google . |
AI Mode on Google Search can also create visual elements, such as images, charts, and simulated videos based on queries through the Generative UI mode. Users can interact with or ask further questions about the data.
According to Google, AI Mode can perform more searches thanks to an upgraded “query fan-out” technique, which breaks down questions into smaller parts to better understand the question's intent and “find new content that might have previously been missed.”
Google didn't hesitate to mention its competitor when describing the Gemini 3 Pro as "replacing empty platitudes with valuable and authentic perspectives."
"It not only answers what you want to hear, but also gives you what you really need to hear," a Google representative added.
 |
Gemini's Dynamic View feature creates a visually interactive experience for certain queries. Image: Google . |
Gemini 3 Pro also provides better support for developers in agent coding and "vibe coding." Tools like Google AI Studio, Vertex AI, and Gemini CLI support Gemini 3, alongside the all-new Antigravity platform, which focuses on agent coding and automation.
Regarding agents, Google says Gemini 3 offers improved support. The model can perform complex processes on behalf of users, such as categorizing and organizing emails, searching for information, and booking travel tickets.
Currently, Google AI Ultra users can experience agent features within the Gemini app using Gemini Agent mode.
Breathe hot air onto your opponent.
Google's new model is seen as a tough challenge for competitors. Sources from the New York Times emphasize that both OpenAI and Anthropic are concerned about being overtaken if Gemini 3 excels in tasks such as coding or image creation.
"Given Google's scale, market share, and pioneering advantage in the search field, Gemini could capture market share and leave OpenAI and other companies lagging behind," said Mike O'Rourke, an analyst at JonesTrading.
He added that the shift to Gemini could spread across the market, raising questions for Oracle and Microsoft, companies that receive billions of dollars from OpenAI to provide cloud computing infrastructure.
 |
Google Search's AI mode can create simulated images based on queries. Image: Google . |
The explosion of AI has also led the tech industry to confront the question of balancing costs and business opportunities. Today's AI systems are primarily used for internet search queries, automating programming.
According to McKinsey & Company, this technology operates in data centers filled with expensive supercomputers, and the industry is expected to spend $7 trillion on these facilities by 2030. Wall Street investors are concerned whether the revenue of Anthropic, OpenAI, Microsoft, and Google will be enough to offset the costs.
"We need high-quality, potential use cases to generate revenue. We haven't reached that point yet," emphasized analyst Ben Bajarin from Creative Strategies.
 |
Gemini 3 Pro's score on the WebDev Arena benchmark conducted by Google. Image: Google . |
Koray Kavukcuoglu, Director of Technology for Google's AI division, said Gemini's accuracy will improve as the company deploys it on Google Search. When combined with the search engine, the model can instantly retrieve and validate information. Of course, this is only effective to a certain extent.
"If people are going to Google just to avoid clicking on blue links, that's not good enough right now," said Manos Koukoumidis, co-founder of Oumi, a startup that helps businesses automate and deploy AI solutions.
To attract more users, Google announced it will offer a free one-year AI Pro package to college students in the US. The company listed over 20 partners who use Gemini 3 to analyze data and record multilingual meetings.
According to Bajarin, these new features could help Google compete with Microsoft and Amazon in the cloud computing market. In October, Google announced that its cloud revenue for the most recent quarter reached $15 billion , primarily driven by AI demand.
Source: https://znews.vn/google-ra-mat-gemini-3-manh-me-hon-post1603987.html
Comment (0)