Google on the 18th (local time) unveiled Gemini 3, its latest artificial intelligence (AI) model equipped with cutting-edge reasoning. Google Chief Executive Officer Sundar Pichai said that day that "Gemini 3 is the model that brings together all of Gemini's capabilities," adding it "has state-of-the-art reasoning power and can turn any idea into reality."
Google signaled its intent to cement its lead in search by introducing Gemini 3 into AI search, outpacing OpenAI, Perplexity, and others.
It has been about eight months since Google introduced a new AI model after releasing Gemini 2.5 in Mar. Reasoning ability as well as coding skills and multimodal (text, image, video, PDF, and other composite inputs) comprehension have improved significantly.
Google said Gemini 3 is designed to grasp an unprecedented level of depth and nuance, such as catching subtle cues in creative ideas or solving complex problems. Pichai said, "In just two years, AI has advanced beyond simply reading text and images to 'reading the room.'"
Google made a bold move by applying Gemini 3 directly to search, its core service. Users can enter a query in the Google search bar and then move to the AI mode tab to use Gemini 3 in a format similar to existing AI chatbots.
Until now, Google had been seen as relatively cautious about bringing AI into search out of concern for self-cannibalization that could eat into revenue in search advertising, its cash cow. However, by applying the new model to search, the company appears to have shifted strategy to a head-on contest by actively using AI in the search institutional sector as well.
Pichai also stressed, "This is the first time we are applying a Gemini model to search from day one of its launch." However, the AI mode in Google Search with Gemini 3 will roll out first in the U.S. market, followed by Korea and other countries in stages.
According to Google, Gemini 3 Pro scored higher than competing models such as OpenAI's GPT-5 on benchmark metrics that evaluate AI model performance. On Humanity's Last Exam (HLE), which assesses Ph.D.-level reasoning, it scored 37.5%, ahead of GPT-5 (25%) and Claude Sonnet 4.5 (13.7%). On the LM Arena leaderboard, which is evaluated directly by users, it scored 1501 points, surpassing Grok 4.1 and the previous Gemini 2.5 Pro.
On Math Arena Apex, which consists of the most difficult problems even among math competition-level questions, it far exceeded the previous top score of 5.21% to achieve 23.4%.
Google cited multimodal performance as a strength of Gemini 3. The new model has evolved into Multimodal AI that provides information in various formats—such as visual layouts generated on the fly to match search intent, interactive tools, and simulations.
Coding performance also improved. On the WebDev Arena leaderboard, a benchmark that measures AI model coding performance, it scored 1487 points, surpassing GPT-5 (1473 points) and Anthropic's Claude Opus 4.1 (1451 points).