This past Tuesday, Google released Gemini 3 – its much-anticipated new flagship natively-multimodal model. It was originally slated for Dec 2025 but prediction markets were heavily in favor of a release this week and hints from leaders like CEO Sundar Pichai and DeepMind head Demis Hassibis suggested it might come sooner than expected. (OpenAI’s quiet release of the well-received but less-impressive GPT-5.1 last week might also have had something to do with it.) The benchmarks for Gemini 3 appear to be significantly outpacing Gemini 2.5 Pro. The model is receiving rave reviews and strongly topping the leaderboards in Text, WebDev, and Vision, boosting Alphabet’s share price up 6.9% on the release. It’s not Google’s only announcement in the past week either – the past 7 days has seen a steady parade of new AI-related releases from Google.

Google is “shipping Gemini at the scale of Google.” On Day 1, Gemini 3 is coming to AI Mode in Google Search (the first time a new model has shown up in Search on Day 1), the Gemini app, AI Studio and Vertex AI (with API access for developers), the Gemini CLI (command-line interface) open-source tool, and a new agentic development platform called Google Antigravity. It will be immediately available to all users in the Gemini app, and to Pro/Ultra subscribers in Search and Gemini CLI. In Search, users will see an improved shopping experience and new “Generative UI” experiences in their results like customized calculators and interactive simulations. Gemini 3 will eventually be available in Search to everyone in the US but with higher limits for Pro/Ultra subscribers. Gemini 3 is also available in 3rd-party platforms, including Cursor, GitHub, JetBrains, Manus, Replit, and others.

The flavor immediately available is Gemini 3 Pro, which sits alongside “Fast” (Gemini 2.5 Flash) in the app. Another version, Gemini 3 Deep Think for more complex problems (which outperforms Gemini 3 Pro on key benchmarks like Humanity’s Last Exam), is currently accessible to safety testers and will eventually be available to Google AI Ultra subscribers.

Examples of what Gemini 3 can do include: Translating handwritten recipes in different languages into a shareable family cookbook; generating interactive flashcards and visualizations from academic papers, video lectures and tutorials; analyzing videos of sports matches to generate a training plan for improvement; turning a napkin sketch into a game or website; teaching how RNA polymerase works; coding a visualization of plasma flow in a tokamak (nuclear fusion); and writing a poem about the physics of fusion.

Google’s new Antigravity coding interface – a swing at Cursor and similar players – offers a full-featured agentic IDE (integrated development environment) that supports multiple models, including Gemini 3 Pro, Claude 4.5, and GPT-OSS 120B. The IDE provides a chat-style agent window side-by-side with a multi-tabbed workspace (which has reported similarities to Google acqui-hire Windsurf’s IDE). In addition to tab completions and in-line commands, agents in Antigravity can work across the editor, terminal, and browser. (An Agent Manager view allows the user to control multiple agents at once.) Notably, Antigravity produces tangible deliverables called “Artifacts” (e.g. implementation plans, task lists, walkthroughs, screenshots, browser recordings) as it works, allowing the user to verify the work is done and see the upcoming tasks. (Antigravity will also provide lists of the model’s actions and tool calls.) Antigravity will also continuously incorporate user feedback provided in the comments that can be added to every surface and Artifact, even while the agent is executing.

The feedback has been almost universally positive. Users are saying Gemini 3 is more consistent and less “spiky”; terse and direct (and less sycophantic) in its responses; capable of being coherently creative; fast with a high “intelligence per second”; and able to deliver excellent frontend detail. One X user noted Gemini 3’s ability to “one-shot” a full-stack app in just 42 min. In evals by code-intelligence startup Sourcegraph/Amp, Gemini 3 outperformed Anthropic’s Claude Sonnet 4.5 by 17 percentage points. (Gemini 3 is also 25%+ cheaper than Claude Sonnet 4.5, though pricier than GPT-5.) Sourcegraph immediately switched its default agent to Gemini 3, saying it was the “first model where we felt compelled to switch Amp’s default smart agent on model release day.” Industry watchers have commented that Gemini 3 is showing “genuine judgment” (although Hassibas believes the industry is still 5-10 years away from “proper full AGI”).

Most of the user complaints are centered on hitting limits quickly with Gemini 3 – even with an Ultra plan. Despite the “generous limits” for Gemini 3 Pro in Antigravity, users are complaining about hitting limits very quickly. Gemini 3 is also only 72% accurate – a significant improvement in the state-of-the-art but still far from perfect.

According to one of Gemini’s co-leads, Oriol Vinyals, Gemini 3’s improvements came from better pre-training and post-training. It contradicts the prior narrative of scaling laws being close to hitting a wall. According to Vinyals, pre-training is still far from hitting a wall and post-training is a “total greenfield.” The training was done on Google’s custom TPUs (Tensor Processing Units), which have been making significant headway in vying with Nvidia’s Blackwell chips. Users have noted that the pre-training appears to have been done on data from 2024, with humorous responses ensuing from Gemini 3 when search is turned off.

A few days after Gemini 3’s launch, Google released Gemini 3 Pro Image, also known as Nano Banana Pro. According to Google, the new image-editing model offers “studio-quality control,” with better visuals and more accurate legible text. It is capable of producing both more realistic images and richer infographics.. Users are calling it the “AI Photoshop of my dreams.” Google is also making it easier for users to detect AI-generated content by uploading an image to Gemini and asking whether it is AI-generated.

Google has become the AI player to beat. The scale of Google Search means its AI Overviews now have 2B monthly users. Even the relatively new Gemini app has 650M monthly users. Gemini also has distribution channels through Google Cloud (where 70%+ of customers use Gemini), Android Auto (which saw Gemini’s rollout begin yesterday), Google Maps (which now has a new Gemini-powered “Know Before You Go” feature), and a large developer community with 13M developers who have already built with Gemini models.

While Google was slower out of the gate to commercialize AI, its investments over the past year to re-architect its products’ backend to be fully on the Gemini stack will continue to pay off. Hassibis is calling DeepMind – acquired in 2014 at a now-bargain price of $500M+ and merged with the Google Brain team in 2023 – the “engine room of Google” that develops the underlying models. Google also has its own custom TPUs for training and Google Cloud for hosting. Given all these assets, Google may be right in believing it has a “very differentiated full-stack approach.”

Sep 19 2025 (3 Shifts): Google’s Gemini edges ahead

Aug 22 2025 (3 Shifts): Google's AI for the home and on the go