Google has released Gemini 3 Pro, described as its most intelligent AI model, alongside Gemini Agent. The release comes a few weeks before the first anniversary of Gemini 2. Gemini 3 Pro delivers state-of-the-art reasoning and class-leading coding performance. It integrates across Google Search, the Gemini app, and enterprise platforms, powering features from AI Overviews to real-time interactive tools. The model becomes available immediately in the Gemini app, Google Search’s AI Mode, and developer platforms.
Gemini 3 Pro sets a new benchmark record on Humanity’s Last Exam, one of the most challenging evaluations for AI systems. It achieves 37.5 percent accuracy, surpassing the prior leader, Grok 4, by 12.1 percentage points. This result relies on the model’s standalone capabilities without external tools such as web search, covering complex, diverse questions across subjects.
On the LMArena leaderboard, Gemini 3 Pro takes the top position with 1,501 points. The score covers performance across multiple evaluation categories and positions it ahead of competing models in real-world user preference metrics.
In the Gemini app, Gemini 3 Pro generates more concise responses with improved formatting. Users access it by selecting “Thinking” from the model picker, available to all users. AI Plus, Pro, and Ultra subscribers receive higher rate limits for extended usage.
Gemini 3 Pro powers Gemini Agent, which extends capabilities from Project Mariner, Google’s web-surfing Chrome AI introduced at the end of the previous year. Gemini Agent performs specific tasks on behalf of users, such as organizing messages or drafting replies in an email inbox, rather than providing only general advice. Users grant it access to Google apps like Gmail and Calendar, allowing interaction with personal data and completion of actions within those environments.
In Google Search’s AI Mode, Gemini 3 Pro launches first for AI Pro and Ultra subscribers via the “Thinking” dropdown. It extends to AI Overviews for the most difficult queries. A new routing algorithm for AI Mode and AI Overviews deploys in coming weeks, directing complex questions to Gemini 3 Pro based on query analysis.
Gemini 3 Pro refines AI Mode’s fan-out search technique, which expands queries into multiple parallel searches. The model conducts additional searches and identifies relevant, credible content that earlier models overlooked. Its multi-modal understanding supports dynamic interfaces in responses. For example, in mortgage loan research, it generates an interactive loan calculator where users input variables like interest rates, loan amounts, and terms to compute payments and amortization schedules directly in the search interface.
Developers and enterprise customers access Gemini 3 Pro through the Gemini API, AI Studio, and Vertex AI. These platforms offer APIs for integration, prompt testing in AI Studio, and scalable deployment via Vertex AI for production workloads.
Google released Antigravity, a new agentic coding application. Antigravity programs autonomously, generates its own tasks to advance projects, and delivers progress reports to users. It handles full development cycles, from initial code generation to iterative improvements and status updates. The application breaks down coding projects into subtasks, such as designing algorithms, implementing functions, and debugging. It reports milestones like completed modules or test coverage percentages.
Google introduced Gemini 3 Deep Think, an enhanced reasoning mode, alongside Gemini 3 Pro. Initial availability goes to safety testers, followed by rollout to AI Ultra subscribers. This mode amplifies reasoning depth for demanding tasks, applying iterative thinking to refine solutions before final output.
Gemini 3 Pro’s deployment covers consumer and professional applications. In the Gemini app, its concise formatting reduces verbosity while maintaining completeness. Integration with Gemini Agent enables proactive execution, handling workflows like email triage through direct API calls to Google services.
AI Mode’s fan-out augmentation with Gemini 3 Pro processes queries with greater precision, executing broader search variants to surface authoritative sources. Dynamic interfaces leverage the model’s ability to interpret text, images, and data jointly, producing tools like financial calculators that respond to user inputs in real time.
