Google Launched Gemini 3 With Major Reasoning Boost, New Coding Platform, and Record Benchmark Scores

Gemini 3 becomes Google’s most capable model, released 18 November 2025

Google launched Gemini 3 and Gemini 3 Pro across the Gemini app, AI Mode, AI Overviews, and Search—the first time a new model powers Search at launch. It succeeds Gemini 2.5 (March 2025) and 2.0 (January 2025). Sundar Pichai said the model better grasps “depth and nuance,” requiring less prompting. Google reports 72% accuracy on its internal benchmark. Gemini 3 Pro appears under the “Thinking” model option and is rolling out globally to users over 18. An enhanced reasoning tier, Gemini 3 Deep Think, will launch for AI Ultra subscribers after further testing.

Record-breaking benchmark performance sets a new industry high

Gemini 3 scored 37.4 on Humanity’s Last Exam, beating GPT-5 Pro’s 31.64, and topped LMArena, dethroning Grok 4.1 Thinking. It also performs at the highest tier on GPQA Diamond and sets a record on MathArena Apex. Google calls it its “most intelligent” and “most factually accurate” model. The company notes that real-world reliability improves when paired with live search, though external experts emphasize ongoing limits of benchmark-only evaluations.

Massive reach: over 650M monthly users and rapid enterprise expansion

The Gemini app now reaches 650M MAU (up from 350M in March 2025). AI Overviews reach 2B monthly users, and 13M developers use Gemini. OpenAI reports 700M weekly ChatGPT users. With Gemini 3 Pro embedded into Search and AI Mode, Google is positioning the model as a default layer for consumer and enterprise tasks. Google also offers Google AI Pro free to U.S. college students, saving about $240/year, to accelerate adoption.

Competitive pressure rises as OpenAI, Anthropic, and xAI escalate releases

Gemini 3 followed OpenAI’s GPT-5.1 launch by under a week and Anthropic’s Claude Sonnet 4.5 by two months. xAI’s Grok 4.1 arrived 24 hours before Gemini 3, marking a rapid release cycle among top labs. Internally, OpenAI and Anthropic staff have expressed concern that Gemini 3 may weaken their position, especially in coding and multimodal tasks. Analysts warn that Google’s scale and distribution could impact partners like Microsoft and Oracle. Reports also indicate Apple may explore a custom Gemini-based model for Siri.

Google Antigravity introduces an agentic, multi-pane coding environment

Released alongside Gemini 3, Google Antigravity is a multi-pane IDE with:

• a prompt window
• a command-line interface
• a browser preview pane

It enables agentic workflows where developers describe apps conceptually. It runs Gemini 3 Pro by default but also supports Claude Sonnet 4.5 and GPT-OSS. Antigravity is available on Windows, macOS, and Linux, expanding Google’s push into agentic, cross-platform coding.

Integrated multimodal features create interactive, generative interfaces

Gemini 3 is natively multimodal, processing text, images, and audio simultaneously. It can produce magazine-style layouts, tables, simulations, interactive calculators, recipe-to-cookbook transformations, flashcards from video lectures, and mixed media outputs for travel, art, and history. In the Gemini app, users interact through Visual Layout or Dynamic View modes. The model also powers generative pages and physics simulations directly inside Search. In Canvas, users can build full-featured programs and documents with enhanced coding tools.

Enterprise, Workspace, and cloud capabilities expand via Vertex AI, AI Studio, and Gemini app

Through Vertex AI, Gemini 3 Pro handles onboarding flows, video and factory analysis, procurement automation, and interactive internal tools. Developers access it via the Gemini API and AI Studio. The model leads WebDev Arena and Terminal-Bench 2.0 for coding tasks. For Workspace, Gemini 3 Pro appears under the “Thinking” model option with admin controls via the Workspace console. It is available across Business, Enterprise, Education, AI Pro for Education, Frontline, Essentials, and Nonprofit tiers, rolling out over 15 days. The Gemini app now includes a redesigned interface with a “My Stuff” section. About two dozen companies are already using Gemini 3 for data analysis and multilingual meeting transcription. Google’s cloud revenue rose by one-third to $15B, driven by demand for AI tools.

Tech giants intensify AI infrastructure spending beyond $380B in 2025

Alphabet, Meta, Microsoft, and Amazon collectively project $380B+ in 2025 AI-related capex. McKinsey & Company estimates global AI compute buildout could reach $7T by 2030, raising sustainability concerns as companies look for business models that can justify runaway compute costs. Gemini 3’s deep integration into Search, Workspace, and cloud workflows places Google at the center of this high-cost infrastructure race.

Google emphasizes reduced sycophancy and more direct, context-aware responses

Demis Hassabis said Gemini 3 aims to deliver “genuine insight,” not flattery. Google positions Gemini 3 Pro as more concise, direct, and less sycophantic, with stronger intent inference and trade-off explanation. Search integration enables real-time retrieval for improved grounding, though analysts caution that AI systems remain far from replacing manual verification. The focus on behavior, agentic planning, and interaction design now sits alongside raw model performance as a strategic pillar.

Why This Matters

Gemini 3 is Google’s strongest push against OpenAI, Anthropic, and xAI, combining record reasoning scores, native multimodality, and agentic tools. Deploying Gemini 3 Pro directly into Search, AI Mode, the Gemini app, and Workspace makes Google’s core products large-scale distribution channels for frontier models. Enterprise adoption grows through Vertex AI, AI Studio, and the Antigravity coding environment.

Its 72% accuracy, rising compute demands, and the industry’s projected $7T data-center buildout by 2030 highlight ongoing economic uncertainty. Yet with 650M+ Gemini app users, 2B AI Overviews, and deep Workspace integration, Gemini 3 expands Google’s role across consumer, enterprise, and developer ecosystems.

This article was drafted with the assistance of generative AI. All facts and details were reviewed and confirmed by an editor prior to publication.