Google Gemini 3 release that tops the list: crushing the competition was praised by Musk, training relies on TPU to show strength

Early this morning.googleThe Strongest Reasoning ModelGemini 3Finally unveiled, a model that encompasses a wide range of native multimodal, reasoning, and Agent capabilities.

The Google DeepMind research team says that theThis is the world's most advanced model for multimodal understanding, Google's most powerful Agent programming and ambient programming modelIt is a rich visualization and deeper interactive experience, built on state-of-the-art reasoning technology.

The model is trained on Google TPUs, supports a context window of 1 million tokens, and is suitable for applications that require the following features: agents, advanced programming, long contexts, multimodal understanding, and algorithm development.

Right off the bat, the Gemini 3 has been butchering almost every review set that's out there.Ranked No. 1 in LMArena's Big Model Arena with 1,501 Elo Score.

Sam Altman, co-founder and CEO of OpenAI, and Elon Musk, founder and CEO of xAI, sent “congratulatory messages” to Google. Altman tweeted “Gemini 3 looks great” and Google CEO Sundar Pichai replied with an emoji.

Musk retweeted a tweet from Google DeepMind CEO Demis Hassabis saying “good job”.

Starting today, Google will deploy Gemini 3 on the following platforms:

For all users of the Gemini app, as well as users of Google AI Pro and Ultra subscription services in the AI mode of search;For developers in the Gemini API, developers of Antigravity, Google's new Agent development platform, and developers of the Gemini CLIfor the Vertex AI platform with Gemini Enterprise Edition.business user.

Additionally, Google will open up Gemini 3's Deep Thinking mode to Google AI Ultra subscribers in the coming weeks, and its currently undergoing a security evaluation.

For the release of the Gemini 3, Pichai believes that this model can beMake any idea the user has come up with a reality.

First, build interactive games and apps in minutes, but also help you learn new knowledge

Let's start by looking at what the Gemini 3 Pro can do.

Gemini 3 can code visualizations of plasma flows in tokamak devices and create poems that capture the physics of fusion.

If users want to learn traditional family cooking, the Gemini 3 can decode and translate handwritten recipes in different languages to create shareable family recipes.

Or if users want to learn a new topic, they can feed Gemini 3 academic papers, long video lectures, or tutorials, and it can generate interactive draw cards, visualizations, or other formats of code to help them master the content.

Gemini 3 also analyzes video of the user's pétanque matches, identifies areas that can be improved, and generates a training plan for overall movement enhancement.

The AI search mode enables Gemini 3 to learn complex subject matter, such as complex knowledge points like the mechanism of action of RNA polymerase with the help of the generative user interface of the AI model in the search function. It's worth noting that this is the first time Google has integrated a new model directly into the AI search function on the first day of the model's release.

Gemini 3 can write retro 3D spaceship games with rich visualization interfaces and interactivity.

The model builds, deconstructs and re-creates fine 3D voxel art through code that can bring a user's imagination to life.

Gemini 3 can create playable sci-fi worlds using shaders.

It can also generate interactive web pages and apps that are rich in more useful elements.

Second, the slaughter of the list of review sets, refresh the ceiling of the ability of large models

Check out the Gemini 3 Pro benchmark results again.

Google's blog mentions that the Gemini 3 Pro was evaluated in a series of benchmarks, including inference, multimodal capabilities, Agent tool usage, multilingual performance and long contexts, with itsFar outperforms the Gemini 2.5 Pro in all major AI benchmarks and ranks #1 in LMArena's Big Model Arena with a 1501 Elo score.

The model exhibitsDoctoral level reasoning skillsIn addition to the highest scores on the Ultimate Human Test (37.51 TP4T without the use of any tools) and the GPQA Diamond Test, the latest top score of 23.41 TP4T was achieved on the MathArena Apex Test.

In addition to text, the Gemini 3 Pro earned 811 TP4T on MMMU-Pro and 87.61 TP4T on Video-MMMU for multimode inference. It also received a state-of-the-art 72.11 TP4T on SimpleQA Verify.

This means that Gemini 3 Pro is able to solve complex problems covering a wide range of topics such as science and math with a high degree of reliability.

Gemini 3 Deep Think and Multimodal Understanding have been updated to help users solve more complex problems. In testing, Gemini 3 Deep Think outperformed Gemini 3 Pro in the Human Ultimate Test (41.01 TP4T without tools) and GPQA Diamond (93.81 TP4T). itachieved 45.1% on ARC-AGI-2 (code execution, ARC award certification), both outperforming Google's own predecessor model, as well as models from OpenAI, Anthropic.

Of the programming capabilities, Gemini 3 is the best ambient programming and Agent programming model that Google has built to date.

The model topped the WebDev Arena leaderboard with a 1487 Elo score. It scored 54.21 TP4T on Terminal-Bench 2.0, which tests the model's ability to use tools, and far outperformed 2.5 Pro on SWE-bench Verified, a benchmark that measures the ability to program Agents.

Developers can build with Gemini 3 in Google AI Studio, Vertex AI, Gemini CLI, and Google Antigravity, Google's new agent development platform. It also supports third-party platforms such as Cursor, GitHub, JetBrains, Manus, Replit, and more.

This means that Gemini 3 can help users with everyday tasks such as booking local services or organizing their inbox.

Third, the new Agent development platform debut, to achieve end-to-end software development automation

Today Google also releasedNew Agent Development Platform Google Antigravity.

With Gemini 3's advanced reasoning, tool usage, and Agent programming capabilities, Google Antigravity transforms AI assistance from a tool in a developer's toolkit to a proactive partner.

While the core of Google Antigravity is still an AI Integrated Development Environment (AI IDE) experience, its Agents have been upgraded with proprietary interfaces and direct access to editors, terminals, and browsers. Today, these Agents can autonomously plan and synchronize complex end-to-end software tasks for developers, while also validating their own code.

In addition to the Gemini 3 Pro, Google AntigravityIt will also incorporate Google's latest Gemini 2.5 computer-use browser model, as well as the image editing model Nano Banana.

Google Antigravity has built an end-to-end Agent workflow for flight tracking applications with the help of Gemini 3. The Agent is able to autonomously plan and write application code and validate its execution through browser-based computer operations.

Finally Google also mentioned that Gemini 3 is its most secure model to date and has undergone the most comprehensive security evaluation of any Google AI model. The model evaluation showed a reduction in sycophantic behavior, increased resistance to instant injections, and enhanced protection against cyberattack abuse.

It's been nearly two years since the Gemini model was released in December 2023: Gemini 1's breakthroughs in native multimodality and long context windows expanded the kinds of information that could be processed as well as the amount of processing; Gemini 2 helps users deal with more complex tasks and ideas, and puts Gemini 2.5 Pro ahead of the rankings in LMArena by more than six months.

Today, Google's Gemini model-based search feature, AI Overviews, now has monthly live users of2 billionThe Gemini app has more than650 million, more than 70% cloud customers use Google AI capabilities, and 13 million developers have built works with their generated models.

Conclusion: Free and Open + Performance Boost! Gemini 3 Stirred up the Big Model Competition!

Google Gemini 3 offers a significant increase in performance over previous generations of models, sensing subtle clues and complex questions in user prompted words, as well as understanding the context and intent behind user requests, allowing users to get the information they need with fewer prompts. Google's blog mentions that in the next new chapter of the Gemini 3 release, they will continue to push the frontiers of Intelligence, Agents, and Personalization to make AI truly accessible to all.