The preview of Google's AI assistant Gemini's third version, 'Gemini 3 Pro,' has been released. Since its launch two years ago, Gemini has been used by 2 billion people monthly, with the application boasting 650 million monthly active users, making it a widely used AI. Gemini 3 is equipped with advanced reasoning capabilities that can understand complex questions or subtle clues within ideas and has been improved to more accurately grasp requests and intentions with minimal questioning. While the first generation two years ago was at the level of analyzing text and images, the third generation has evolved to read context and mood.


Gemini 3 is available to free users under the name 'Thinking Mode,' while 'Google AI Ultra' subscribers can try the enhanced version, Gemini 3 Deep Think, first. The AI mode provided when inputting into the existing Google search engine now operates with Gemini 3 search and is applied to Google's 'Gemini' app, 'AI Studio,' the developer tool 'Vertex AI,' and the AI agent development platform 'Google Antigravity.'

Gemini 3 focuses on better understanding and autonomously processing information

Google deployed its proprietary AI accelerator, TPU v5, among other hardware, in the development of Gemini 3, but has not disclosed exactly how much data was used. It is known to be a completely new model rather than a high-performance version of the previous generation, having secured data from web crawling, licensed content, and more.

Google aims to maintain its influence in the AI-based next-generation search engine market through Gemini 3. The AI mode provided with basic searches applies a generative UI that instantly configures a visual layout according to the user's search intent. For example, if searching for images, it provides an image-centric layout; if analyzing papers or focusing on text, it offers a corresponding screen configuration.


The multimodal reasoning function, which comprehensively recognizes various types of data such as text, images, and audio, has been further enhanced. For instance, if a photo of a recipe written in multiple languages by hand is uploaded, the AI automatically converts it into words, translates it comprehensively, and creates a contextually appropriate recipe. If the original text of a paper is input and a request is made to visualize it with a 3D interactive image, it generates a video to explain it. The context that can be processed at once has increased to a maximum of 1 million tokens, allowing for the processing of large-scale coding tasks at once or the immediate input of long content such as books or papers.

AI agents, which have been a hot topic in the industry since last year, are also enhanced starting with Gemini 3. Google is launching the new agent development platform 'Google Antigravity' to assist in not only coding collaboration for developers but also task processing. Antigravity helps with AI agent development tasks based on the enhanced performance of Gemini 3 and has a much wider working area and authority than before, including editors, terminals, and browsers. It also provides a function to self-verify completed code.

Still the highest level of performance, with an even higher 'Deep Think' mode

Performance has also improved dramatically. In the GPQA Diamond Test, composed of graduate-level past exam questions, Gemini 3 Pro recorded 91.9% accuracy compared to 86.4% for the existing 2.5 Pro and 88.1% for GPT-5.1. In the American Invitational Mathematics Examination (AIME) 2025, it recorded a 95% correct rate compared to 94% for GPT-5.1 and 87% for Claude Sonnet 4.5. In Humanity’s Last Exam, composed of over 3,000 questions across more than 100 fields such as mathematics, humanities, and natural sciences, it recorded 37.5% accuracy compared to 26.5% for GPT-5.1 and 13.7% for Sonnet 4.5. Combining external search and code tools can increase accuracy to a maximum of 45.8%.

Gemini 3 Pro has risen to the top of the LMArena leaderboard, which evaluates AI performance by field. The list includes a total of 270 representative AIs, with Grok 4.1-Thinking, Grok 4.1, Claude Sonnet 4.5, Gemini 2.5 Pro, and GPT-5.1 versions following Gemini 3 Pro in order.


Meanwhile, better results can be achieved with the advanced reasoning mode, Deep Think. Deep Think processes calculations for a longer time to increase accuracy, raising the accuracy in Humanity’s Last Exam to 41%. Considering that it is 37.5% in normal mode, the improvement is quite significant. It also achieved a record 45.1% in the ARC-AGI 2 puzzle test, which requires human-level flexible thinking. Existing models remained at about 10% to 15%.


The usage price of 'Google AI Pro,' which includes the Gemini 3 subscription, remains the same at KRW 29,000 per month. The usage price of 'Google AI Ultra,' which includes advanced user access and early access to Gemini 3 Deep Think, also remains at KRW 360,000, excluding promotions.

The API integration cost required for external service utilization has increased significantly. The API integration cost for Gemini 2.5 Pro was USD 1.25 (approximately KRW 1,836) per 1 million tokens for input and USD 10 (approximately KRW 14,690) for output. The Gemini 3 Pro preview API currently costs USD 2 (approximately KRW 2,938) per 1 million tokens for input and USD 12 (approximately KRW 17,632) for output, with input costs increasing by 60% and output costs by 20%. Instead of adjusting usage by limiting the number of tokens, this version allows cost adjustment by setting the intensity of inference to Low or High.

An AI that captures both usability and performance, with expectations for derivative versions

The industry's response is enthusiastic. OpenAI's GPT-5 was re-released less than a week after its launch, with plans to further develop future versions. In contrast, Gemini 3 Pro is already being talked about for its excellence among developers and has no complaints among actual users. Most free and paid users report no significant inconvenience. Considering that more impactful derivative versions such as Flash and Nano have not yet been released, and that it could influence future performance improvements in Veo, Imagen, Med-Gemini, and others, the future looks promising. It is predicted that Google will win this year's AI trophy.

IT Donga Reporter Nam Si-hyun (sh@itdonga.com)

Popular News