Google Unveils Gemini 3: A Cutting-Edge Multimodal Agent Model More Powerful Than Ever
After weeks of speculation and leaks, Google officially unveils Gemini 3. This suite of closed models is designed as the most ambitious iteration from Google since the introduction of Gemini in 2023. The update features a thoroughly reinvented multimodal engine, significantly enhanced agent capabilities, and complete integration within the Google ecosystem.
A Major Launch: Gemini 3 Arrives Everywhere, All at Once
Unlike previous fragmented launches, Gemini 3 is being launched simultaneously across Google Search (AI Mode), the Gemini app, Google AI Studio, Vertex AI, Gemini CLI, and even Antigravity, Google’s brand-new agent-first environment.
The company describes this as its most coordinated rollout to date, made possible by complete control over the supply chain: in-house TPUs, data centers, consumer products, and developer tools.
With 650 million active monthly users on the Gemini app and 2 billion interactions with AI Overviews in Search, Google is targeting a massive rollout right from the start.
A Shift Toward Agent-based AI
At the core of Gemini 3 is a new philosophy. Google aims not just to generate text or interpret images; the goal is now to plan, navigate, execute, and coordinate tools within a complete software environment.
Gemini 3 comes in several versions:
- Gemini 3 Pro, the flagship model
- Gemini 3 Deep Think, an intensive reasoning mode
- Generative interface models (Visual Layout, Dynamic View)
- Gemini Agent, capable of orchestrating complex workflows
- The Gemini 3 engine integrated with Antigravity, aimed at developers
The ambition is clear: to transform the conversational assistant into a true digital operator, capable of action.
Gemini 3 Pro: Spectacular Gains in Reasoning, Coding, and Multimodality
The benchmarks provided by Google are impressive. Gemini 3 Pro outperforms Gemini 2.5 Pro on nearly all fronts:
- Reasoning: Ranked first on the LMArena leaderboard with a preliminary Elo of 1501, ahead of Grok-4.1-thinking, Grok-4.1, and Claude Sonnet/Opus.
- Mathematics: 95% on AIME 2025 without tools, 100% with code execution.
- Vision & video: +13 points on MMMU-Pro, spectacular progress on ScreenSpot-Pro (11% → 72.7%).
- Code & tools: LiveCodeBench Pro (+664 points), SWE-Bench Verified (59.6% → 76.2%).
- Long context: 26.3% at 1M tokens, a major rating boost for latency-sensitive tasks.
- Planning: Vending-Bench x10 compared to the previous model.
In practice, Gemini 3 becomes a model capable of long reflections, screen navigation, and stable use of tools—three historical challenges of AI models.
Generative Interfaces: Google Aims to Go Beyond Text
Gemini 3 introduces Visual Layouts and Dynamic Views, real-time generated interfaces including:
- Magazine pages,
- Interactive simulators,
- Graphs,
- Complete UI modules.
In Google Search, the experience becomes more visual and immersive: simulations, tables, galleries, and interactive tools are integrated directly into the responses.
Google presents this as a step closer to a search engine where AI doesn’t just explain, it builds.
Gemini Agent: The AI That Acts, Not Just Responds
Initially available for AI Ultra subscribers, Gemini Agent can sort emails, draft replies, plan trips, prepare documents, browse the web, and orchestrate multi-step tasks.
This all happens with user validation—a compromise between automation and controllability.
Antigravity: The Development Environment Merging AI, Code, and Navigation
Google also introduces Antigravity, a kind of super-agent IDE where Gemini 3 can write code, prototype interfaces, test, run, fix, manipulate browser UI, generate reports, and create internal tools.
The model can even interpret mouse movements, visual annotations, and multi-window mockups to guide its actions. A true breakthrough for automated computer use.
API: An Ambitious Pricing Strategy
Gemini 3 Pro is priced at:
- $2/M tokens input
- $12/M tokens output
- Up to 200k tokens,
- With nearly double the rate beyond that.
Facing Chinese alternatives (Qwen, ERNIE) that are much cheaper, Google positions itself in the high-end segment—a risky bet for startup adoption, but one they are committed to for a model presented as premium.
A hype rarely seen for a Google model
A Veiled Jab at OpenAI
Google does not hide its confidence: Gemini 3 Pro is claimed to be “more direct, more relevant, and less flattering”—a clear reference to recent adjustments made to ChatGPT to reduce its sycophancy.
The company also asserts that it provides “more factual” and “more useful” responses, a territory where OpenAI has faltered with the complex launch of GPT-5.
A Strategic Advance for Google
With Gemini 3, Google aims to prove three things:
- Its hardware (TPU) is ready for frontier AI
- Its models are competitive once again
- Its consumer ecosystem can serve as a massive launchpad
Beyond the benchmarks, the company hopes that useful AI deployed in Search, Android, Workspace, and Gemini will prevail over a raw race for power.
