Google launches Gemini 3 with big reasoning leaps

For the past two years, big AI announcements followed an exhausting script: drop a slick research paper, tease a cherry-picked video demo, and leave everyone stuck on a waitlist. Google just broke that habit by releasing its most advanced system to date directly into the products people open every single morning.

Google launched Gemini 3, headlined by the Gemini 3 Pro preview. Alongside the base model, Google revealed Gemini 3 Deep Think, an enhanced reasoning engine, and introduced Google Antigravity, a brand new agentic developer platform.

Benchmarks fall while responses turn interactive

On the scorecard, Gemini 3 Pro took the crown across major evaluations. It topped LMArena with an unprecedented 1501 Elo and scored 1487 on WebDev Arena. On Humanity's Last Exam, an evaluation designed to stump advanced systems without web help, it scored 37.5% without tools, while hitting 91.9% on GPQA Diamond and 76.2% on SWE-bench Verified.

The specialized Deep Think mode pushed those limits even higher, scoring 41.0% on Humanity's Last Exam and 45.1% on ARC-AGI-2 with code execution.

Yet the most noticeable shift is what Google calls generative interfaces. Instead of returning paragraphs of text, Gemini 3 writes and renders working software on the fly.

In Search AI Mode and the Gemini app, querying complex physics like the three-body problem no longer yields a wall of text. It spawns a functional gravity simulator where you adjust variables on a canvas. Asking for mortgage advice builds an interactive calculator right inside the response. Travel itineraries arrive styled like interactive digital magazines rather than plain bullet points.

Shifting developers from copilots to agents

For builders, Gemini 3 signals a move toward autonomous software creation.

Google launched Antigravity, a free public preview available on macOS, Windows, and Linux. While it looks like an AI code editor, agents inside the app have direct access to the editor, terminal, and browser. They write code, execute commands, and run browser-based tests to debug their own work before asking for human approval.

Through Google AI Studio and Vertex AI, developers can access the 1 million token context model at $2 per million input tokens and $12 per million output tokens for prompts under 200k tokens. It also plugged directly into Cursor, GitHub, JetBrains, Replit, and Manus on day one.

For power users, Google is testing Gemini Agent with Ultra subscribers. Powered by Gemini 3, it reads context across Gmail and Google Calendar to orchestrate real-world errands, such as pulling flight times to reserve a rental car under a target budget, pausing only to seek confirmation before checkout.

Why it matters

Google flexed its unique superpower here: massive distribution. While frontier rivals often launch models in isolated playgrounds, Google shipped Gemini 3 into Google Search, Gemini, and developer workflows on day one. It is a calculated move to prove that state-of-the-art intelligence belongs directly inside mass-market utilities.

For ordinary users, generative UI is a massive upgrade over conversational chat. When exploring complex topics or comparing products, interacting with a live widget is far faster than typing five follow-up prompts. For developers, Antigravity and vibe coding make clear that building software is shifting from typing syntax to managing autonomous swarms of agents.

There are clear trade-offs, though. Generating custom code and user interfaces on the fly demands intense compute, which can easily trigger latency spikes. If a dynamically generated simulator miscalculates physics or an agent misinterprets flight confirmations, users might quickly lose trust. Google solved the raw intelligence hurdle, but keeping these instant apps stable and fast at billions-of-users scale will be the real test.