Gemini 3: Google's Most Intelligent Model Yet
Google's Gemini 3 launch shows where frontier modeling went next: reasoning depth, agentic coding, multimodal scores, and a same-day rollout across Search.
The easy assumption is that a frontier model launch means higher benchmark scores and little else changes. With Gemini 3, Google shipped the model into Search on launch day for the first time — and weighted behavior, like dropping flattery and needing fewer prompts, as heavily as the scores.
Source
A new era of intelligence with Gemini 3
By: Google (Sundar Pichai, Demis Hassabis, and Koray Kavukcuoglu, on behalf of the Gemini team)
Published: Nov 18 2025
Type: Product announcement
This is the reference point for where Google’s model line stood at the end of 2025: the full generation story from Gemini 1 to 3, the benchmark case, and the product-surface rollout collected in one post. It’s also a clean example of a launch that spends as much time on what the model is like to work with as on what it scores.
What Gemini 3 changed beyond the scores
The intro builds the generation story in three steps. Gemini 1 brought native multimodality — one model handling text, images, audio, video, and code — plus the long context window. Gemini 2 laid the agentic foundation and pushed reasoning, with 2.5 Pro topping the LMArena leaderboard for over six months. Gemini 3’s stated job is to combine all of it into Google’s most intelligent model. The usage numbers explain the stakes: AI Overviews at 2 billion users a month, the Gemini app past 650 million monthly users, more than 70% of Google Cloud customers using its AI, and 13 million developers building with its generative models.
The release itself is Gemini 3 Pro, in preview, described by Demis Hassabis and Koray Kavukcuoglu as the best model in the world for multimodal understanding and Google’s most powerful agentic and “vibe coding” model — vibe coding being the practice of building software by describing what you want rather than writing the code yourself. Alongside it sits Gemini 3 Deep Think, an enhanced reasoning mode held back for extra safety review before reaching Google AI Ultra subscribers.
The benchmark paragraph is dense but worth decoding. LMArena ranks models by pairwise human preference using an Elo rating — the chess-style scoring — and Gemini 3 Pro topped it with 1501 Elo. It scored 37.5% on Humanity’s Last Exam without tools, 91.9% on GPQA Diamond for graduate-level science, and a state-of-the-art 23.4% on MathArena Apex. On multimodal reasoning it posted 81% on MMMU-Pro and 87.6% on Video-MMMU, plus 72.1% on SimpleQA Verified, a factual-accuracy check. Deep Think goes further: 41.0% on Humanity’s Last Exam, 93.8% on GPQA Diamond, and 45.1% on ARC-AGI-2, a benchmark of novel puzzles built to resist memorization. Google also makes a point of the model’s tone — responses that trade “cliché and flattery for genuine insight — telling you what you need to hear, not just what you want to hear.”
The post then organizes what you can do into learn, build, and plan. For learning, the claim is multimodal synthesis over a 1M-token context: deciphering handwritten recipes in different languages into a shareable family cookbook, turning academic papers and long video lectures into interactive flashcards and visualizations, or analyzing video of your pickleball match into a concrete training plan. In Search, AI Mode uses Gemini 3 to generate what Google calls generative UI — immersive layouts, interactive tools, and simulations created on the fly for a query, with an RNA polymerase explainer as the showcase.
For building, the numbers move to agentic coding: 1487 Elo on WebDev Arena, 54.2% on Terminal-Bench 2.0 (which tests operating a computer through a terminal), and 76.2% on SWE-bench Verified, the standard measure of coding agents. The structural news is Google Antigravity, a new agentic development platform where agents get a dedicated surface with direct access to the editor, terminal, and browser, and can plan and execute end-to-end software tasks while validating their own code — flanked by the Gemini 2.5 Computer Use model for browser control and the Nano Banana image model. Gemini 3 also reached third-party tools on day one, including Cursor, GitHub, JetBrains, Manus, and Replit.
The planning claim is the most unusual. Gemini 3 tops Vending-Bench 2, a benchmark where a model runs a simulated vending-machine business, by holding consistent tool use and decision-making across a full simulated year of operation without drifting off task. That is a claim about persistence, not intelligence — the failure mode it targets is the agent that forgets the goal twenty steps in. Google ties it to everyday delegation: booking local services or organizing an inbox as multi-step workflows under your control, with a Gemini Agent already live for Ultra subscribers.
The safety section claims the most comprehensive evaluation set of any Google AI model, and its items are behavioral: reduced sycophancy (the model telling you what you want to hear), stronger resistance to prompt injection (malicious instructions hidden in content the model reads), better protection against cyber misuse, plus external assessments and early access for the UK’s AI Safety Institute. The tell is where the flattery fix sits in the post — under safety, not style.
The rollout list is the quiet headline. Gemini 3 landed in the Gemini app for everyone; in AI Mode in Search for AI Pro and Ultra subscribers — the first time Google shipped Gemini in Search on day one; in AI Studio, Gemini CLI, Vertex AI, and Antigravity for developers; and in Gemini Enterprise for organizations, with more models in the family promised soon.
My take
Read as a whole, this is a distribution announcement wearing a benchmark announcement’s clothes. The scores get one paragraph; the surfaces get several. When a single model ships simultaneously into an app with 650 million monthly users, the default search experience, a CLI, and a brand-new developer platform, model quality stops being a standalone product and becomes an infrastructure property. The interesting question shifts from “is it smarter?” to “what happens when every surface quietly swaps its model on the same Tuesday?” Two smaller tells: staging Deep Think behind safety testers treats raw reasoning depth as a capability needing gated rollout — a more honest frame for “state-of-the-art reasoning” than any single Elo number. And filing reduced sycophancy under safety is, I think, correct and overdue: a model that flatters you is making errors you can’t see, which is a reliability problem before it is a personality problem.
Where this is useful
- Evaluating frontier models for agentic coding — the Terminal-Bench, SWE-bench Verified, and WebDev Arena numbers here are a comparison set to hold any vendor against.
- Deciding whether extra-reasoning tiers like Deep Think are worth paying for on hard analysis work, versus standard inference for routine tasks.
- Understanding the Antigravity bet — agents with editor, terminal, and browser access — before adopting any agentic development platform.
- Treating the scores with care: every number is Google evaluating its own model, some of it on benchmarks it helped popularize.
Apply this: Next time a frontier model launches, count where the announcement spends its words — benchmark claims versus rollout surfaces. The ratio tells you what the vendor believes it is actually competing on.
What I want to remember
Remember: Frontier model launches are distribution events now. The scores move at the margin; the durable change is which surfaces the model reaches on day one — and behavioral fixes like reduced flattery count as safety work, not polish.