The company made the announcement on Thursday, describing the new system as a big step forward from Gemini 3, which it introduced in November 2025. It is already rolling out the preview across consumer, developer and enterprise products.

In a statement, Google said: “Last week, we released a major update to Gemini 3 Deep Think to solve modern challenges across science, research and engineering. 

“Today, we’re releasing the upgraded core intelligence that makes those breakthroughs possible: Gemini 3.1 Pro. We are shipping 3.1 Pro across our consumer and developer products to bring this progress in intelligence to your everyday applications.”

Developers can now access Gemini 3.1 Pro in preview through the Gemini API in Google AI Studio, Gemini CLI, Google Antigravity and Android Studio.

Enterprise customers can use it in Vertex AI and Gemini Enterprise. Consumers are getting it through the Gemini app and NotebookLM.

Google says the new model improves core reasoning. It reports that Gemini 3.1 Pro scored 77.1 per cent on ARC-AGI-2, a benchmark designed to test how well a system can solve entirely new logic patterns. According to the company, that score is more than double the reasoning performance of Gemini 3 Pro.

On another benchmark known as Humanity’s Last Exam, Google says Gemini 3.1 Pro outperformed Anthropic’s Claude Opus 4.6 and OpenAI’s GPT-5.2. Independent testing platform APEX-Agents also ranked the model first for professional task performance.

Brendan Foody, chief executive of Mercor, which runs the APEX system, wrote on social media: “Gemini 3.1 Pro is now at the top of the APEX-Agents leaderboard,” adding that the model’s results show “how quickly agents are improving at real knowledge work.”

Google describes Gemini 3.1 Pro as a stronger base for complex problem-solving. It says the system can support tasks that require detailed reasoning, data synthesis and multi-step workflows.

The company appears to be focusing on tools that go beyond simple text generation and can handle layered instructions.

The update comes just three months after the release of Gemini 3. In that same period, competitors including OpenAI and Anthropic have also introduced new models, intensifying competition among large technology firms.

Google DeepMind has published a model card for Gemini 3.1 Pro outlining its intended uses, limits and safety measures.

The company says the model is natively multimodal, meaning it can process text, images and other forms of input. While that expands its use cases, it also leads to safety and content moderation questions.

For now, Google says the preview will allow it to test improvements further before a general release. Higher usage limits are being introduced for users on the Google AI Pro and Ultra plans, while NotebookLM access is limited to those subscribers.

The company ended its announcement with a brief note to users: “We can’t wait to see what you build and discover with it.”

This is one of the first big exits since Mark Zuckerberg reorganised Meta’s AI division under Superintelligence Labs.

LeCun, a Turing Award laureate and one of the pioneers of deep learning, is reportedly in early discussions with investors to raise funds for his new venture, which will centre on developing “world models”, AI systems designed to simulate and understand the real world more deeply. 

This approach aims to create machines that can learn and reason with a closer resemblance to human cognition.

His departure comes at a time when Meta is enhancing its drive to compete with OpenAI, Google DeepMind, and Anthropic in the superintelligent systems space.

Mark Zuckerberg recently consolidated Meta’s AI research under Superintelligence Labs, placing Alexandr Wang, the former CEO of data-labelling firm Scale AI, in charge of the division.

This reorganisation meant that LeCun, who had long reported to Chief Product Officer Chris Cox, was reassigned to report directly to Wang, a move that, according to sources cited by the Financial Times, may have influenced his decision to leave.

LeCun’s planned exit reveals both a generational change in leadership and a potential divergence in vision between academic research and the dynamic, product-driven approach of Meta AI.

Meta, the parent company of Facebook and Instagram, has yet to comment publicly on the reports. LeCun, too, has not issued an official statement.

Since joining Meta (then Facebook) in 2013, Yann LeCun has helped in promoting convolutional neural networks (CNNs) and self-supervised learning, two techniques that underpin today’s large-scale AI systems. 

His next startup could represent a return to the more exploratory and research-oriented roots that first defined his career.

The platform, built on Google’s Gemini models, aims to strengthen the tech giant’s competitive edge in enterprise AI, as Microsoft, OpenAI, and Anthropic, among others, are not holding back.

Gemini Enterprise functions as a conversational system that enables employees to interact directly with company data, documents, and applications. The goal, Google says, is to make AI a core part of every workflow, not just an add-on. “We’re introducing Gemini Enterprise, designed to bring the full power of Google’s AI to every employee, for every workflow,” the company said during the launch.

The announcement comes as Google Cloud continues to expand, having surpassed a $50 billion annual revenue run rate in the second quarter of 2025. According to the company, about 65% of its cloud customers already use Google’s AI products, including nine of the world’s top ten AI labs.

Gemini Enterprise builds upon Google’s full-stack AI strategy, combining the strength of its infrastructure, foundational models, and research divisions such as Google DeepMind. The platform is powered by a multi-layer system, from the company’s purpose-built Tensor Processing Units (TPUs) and Nvidia GPUs, to its world-leading Gemini models, which have consistently topped global performance benchmarks.

Google describes Gemini Enterprise as “the new front door for AI in the workplace.” It features pre-built AI agents capable of conducting deep research, generating data insights, and automating complex workflows. Through a no-code interface, companies can also create and deploy their own agents, tailored to their specific operations.

Early adopters of the platform include Gap, Klarna, and Figma. Klarna, for instance, is using Gemini’s generative tools to produce personalised lookbooks that have boosted customer orders by 50%. 

In the healthcare sector, HCA Healthcare has deployed a “Gemini-powered Nurse Handoff solution” that simplifies patient information transfers between shifts, a move expected to save millions of hours annually.

Google is also using its own technology internally. Nearly half of all new code at the company is now generated by AI and reviewed by engineers, significantly accelerating development cycles.

The launch further reveals Google’s vision for a connected AI ecosystem. Gemini Enterprise securely links with data across multiple platforms, including Google Workspace, Microsoft 365, Salesforce, and SAP, ensuring enterprise-wide access without compromising governance or security.

To encourage adoption, Google is rolling out Google Skills, a free training platform that will teach users how to build and deploy agents within Gemini Enterprise. The company also introduced Delta, a team of expert AI engineers who will work directly with clients to deploy advanced solutions.

With Gemini Enterprise, Google is embedding AI at the very foundation of how organisations work. “AI is presenting a once-in-a-generation opportunity to transform how you work, how you run your business, and what you build for your customers,” the company stated.

The rollout, announced at its Made on YouTube live event on Tuesday, challenges TikTok and Instagram Reels in the race for creator attention.

One of the most anticipated updates is Veo 3 Fast, a streamlined version of Google DeepMind’s text-to-video model. Unlike the full Veo 3, this variant produces clips at 480p with minimal delay and, for the first time, sound. 

The feature opens up new possibilities where creators can animate still images, transfer motion from one subject to another, and apply unique visual styles such as origami or pop art, all through text prompts.

YouTube is also expanding how creators remix content. A new Speech to Song tool, powered by Google’s music model Lyria 2, can transform lines of dialogue into soundtracks. “As the world’s largest creative playground, YouTube is where trends are born and where you can draw inspiration from. Imagine hearing a line of dialogue that sparks an idea—a funny phrase, a memorable quote, or a one-of-a-kind sound—and you want to remix it into a new sound,” said Dina Berrada, YouTube’s director of Product for Shorts and Generative AI Creation. 

The platform says creators can further personalise these soundtracks with moods like chill, danceable, or fun. Testing has already begun, with broader availability in the United States expected soon.

Another addition, Edit with AI, tackles the challenge of turning raw clips into polished drafts. The feature automatically selects standout moments from footage, adds music, transitions, and even voiceovers in English or Hindi that respond to the video’s action. 

While still experimental, YouTube is trialling it on Shorts and its Create app, with plans to extend to more regions in the coming months.

For now, these tools are available in the United States, United Kingdom, Canada, Australia, and New Zealand, with global expansion projected for 2026. To maintain transparency, every AI-generated Short will carry SynthID watermarks and content labels.

YouTube is lowering the limitations to professional-level content creation. In embedding Google’s most advanced generative AI models directly into the Shorts workflow, the company is arming mobile-first creators with tools to innovate, compete, and keep pace with the speedy evolution of short-form content culture.

“Today was my last day at xAI, the company that I helped start with Elon Musk in 2023,” Igor Babuschkin announced on Wednesday in a post on X. “I still remember the day I first met Elon, we talked for hours about AI and what the future might hold. We both felt that a new AI company with a different kind of mission was needed.”

The engineer, who previously worked on AlphaStar at Google DeepMind and held research roles at OpenAI, was key in building xAI into a major player in the sector in less than two years. 

He oversaw infrastructure, product and applied AI projects, and played a central role in constructing the company’s Memphis, Tennessee supercomputer, completed in just 120 days, despite industry veterans calling the goal “impossible.”

xAI’s Legal Head Robert Keele Resigns Citing Family, Differences with Musk

Babuschkin recalls one defining moment during that build, when a late-night debugging session with Musk led to the discovery of a critical BIOS setting error. “I learned 2 priceless lessons from Elon: #1 be fearless in rolling up your sleeves to personally dig into technical problems, #2 have a maniacal sense of urgency,” he wrote.

His time at xAI comprised of rapid technical achievement but also controversy. The company’s Grok chatbot has faced repeated public backlash, ranging from inserting extremist conspiracy theories into responses to antisemitic rants, to enabling the creation of AI-generated nude videos of public figures such as Taylor Swift. Grok was even suspended from X earlier this year for inflammatory political claims.

Environmental concerns have also followed xAI’s expansion. The Memphis supercluster, dubbed “Colossus”, is powered by 35 methane gas turbines. Local residents and advocacy groups, including the NAACP, have filed appeals over increased air pollution in majority-Black neighbourhoods. 

A University of Tennessee study reported nitrogen dioxide levels near the site had risen by 79%, with local asthma cases spiking in the nearby Boxtown community.

Despite these challenges, Babuschkin describes his departure with affection. “As I drive away today, I feel like a proud parent, driving away after sending their kid away to college,” he said. “My heart is brimming with tears of joy, rooting for the company as it grows and matures.”

His new firm, Babuschkin Ventures, will fund startups developing AI systems aligned with human values, and will support research into AI safety. He says the inspiration came from a recent dinner with Max Tegmark, founder of the Future of Life Institute, where they discussed how to ensure “our children can flourish” in a world with advanced AI.

“The singularity is near, but humanity’s future is bright,” Babuschkin stated, noting his belief that safe, beneficial AI could help unlock solutions to some of the world’s most profound scientific problems, including quantum gravity and the Riemann hypothesis.

Unlike previous initiatives limited to either philanthropy or regional support, this programme is structured to support startups beyond funding.

We’re looking at a full-stack support model that includes early access to AI models like Gemini, Imagen, and Veo; technical assistance from Google DeepMind and Google Labs; Google Cloud credits; and in some cases, direct equity investment from Google.

Startups across different growth stages, whether at seed or scaling globally, can now apply to benefit from this hands-on support, starting 12 May.

What sets this apart is the deliberate involvement of Google’s core AI engineering, product, and commercial teams. Startups won’t just be customers or partners, they’ll be collaborators.

So far, a few names have already gained from this early-stage programme. Toonsutra, a digital comic platform in India, is using Gemini to translate content across the country’s complex language terrain.

Another, Viggle, a meme-making platform, has started testing video content creation using Gemini, Imagen and Veo. Rooms, a 3D interaction platform, is also part of the pilot group, experimenting with new content formats using Google’s tools.

Here’s how Google laid it out:

“Startups working with the AI Futures Fund will receive the following:

• Early access to select AI models from Google DeepMind: This includes some of our most advanced AI models like Gemini, Imagen for image generation and Veo for video generation.
• Collaboration with Google experts: Startups get hands-on support from researchers, engineers, product managers, designers and go-to-market specialists from Google DeepMind and Google Labs.
• Google Cloud credits and support: The program provides Cloud credits and dedicated technical support from our Cloud experts to help build, train and scale AI-powered products.
• Equity investment: Select startups get the opportunity to seek direct investment from Google to fuel growth and scale AI development.”

The launch of the AI Futures Fund is just one part of Google’s vision to shape AI’s growth through structured, high-impact partnerships.

It follows a string of high-value initiatives, including a $120 million Global AI Opportunity Fund and two $20 million commitments: one to nonprofits working on generative AI and another to scientists and researchers tackling global problems with artificial intelligence.

Google is betting big on startups that are willing to build AI.

Startups interested in applying for the Google AI Futures Fund can begin the process now. The fund is open to companies in any region where Gemini is available.

The announcement came directly from Omar Sanseviero, a developer relations engineer at Google DeepMind, who confirmed the figures on X over the weekend. 

In addition to the downloads, more than 70,000 custom variations of Gemma have already been built on Hugging Face, a widely used platform among developers.

Gemma, which launched in February 2024, is part of Google’s initiative to compete with other open model providers, especially Meta’s Llama series. 

The latest versions of Gemma can work with both images and text, a capability known as being “multimodal.” They also support more than 100 languages and include versions optimised for tasks like drug research.

However, when placed side-by-side with Meta’s Llama, Gemma’s traction seems to be on a moderate level. Llama passed 1.2 billion downloads in April 2025, nearly ten times what Gemma has achieved. In a space where momentum matters, such a gap is hard to ignore.

But volume alone doesn’t determine usefulness. What’s more important is what developers can actually do with these models. And that’s where the trouble starts. 

Both Gemma and Llama have been flagged by developers for having restrictive and unclear licensing. These are not the standard open licenses many in the community are used to. Some have argued that the vague terms make it risky to deploy the models in commercial settings.

That has led to hesitation, particularly among startups and businesses that can’t afford legal ambiguity. I’ve spoken with teams who say they’ve skipped over both Gemma and Llama entirely for this reason. The concern isn’t about capability, but clarity.

So while Google celebrates 150 million downloads, there are issues about what those downloads actually represent. Are developers using the models in real products? Or are they just experimenting, unsure whether they can actually go to market with them?

Disney, known for its entertainment technology, aims to be one of the first companies to integrate Newton into its next-generation robotic characters. Among them are the BDX droids, inspired by Star Wars, which are expected to debut at various Disney theme parks next year. 

A live demonstration of one of these droids took place during Nvidia CEO Jensen Huang’s keynote at GTC 2025, where the robot waddled onto the stage.

The Newton engine aims to bridge a challenge in robotics—the difference between simulated environments and real-world physics. Robots usually find it difficult with tasks requiring fine motor skills or adaptability to unpredictable surroundings. 

Newton seeks to change that by providing a physics simulation tool that allows developers to program robots to interact with soft and deformable objects, such as food, fabric, and sand.

Kyle Laughlin, senior vice president of Walt Disney Imagineering, said, “The BDX droids are just the beginning. We’re committed to bringing more characters to life in ways the world hasn’t seen before, and this collaboration with Disney Research, Nvidia, and Google DeepMind is a key part of that vision.”

Nvidia intends to release an open-source version of Newton later this year, allowing developers worldwide to contribute to its evolution. The engine is designed to work seamlessly with Google DeepMind’s MuJoCo, a physics simulator already widely used in robotics research. 

With MuJoCo’s ability to model multi-joint movements and contact-rich environments, the addition of Newton is expected to improve robotic control and precision.

DeepMind’s involvement brings another layer of advancement to the project. Its recent launch of MuJoCo-Warp—a version of the simulator optimised for Nvidia’s GPU acceleration—has demonstrated speed improvements of over 70x for humanoid robot simulations and a 100x boost for in-hand manipulation tasks. By integrating this technology into Newton, Nvidia aims to enhance performance while maintaining high levels of adaptability.

Beyond its applications in entertainment, Newton has the potential to bolster industries that rely on robotics, including manufacturing, healthcare, and logistics. The physics engine is built on Nvidia Warp, a CUDA-based acceleration library that allows developers to create highly efficient, real-time simulations. 

It is also designed with extensibility in mind, meaning it can evolve alongside advancements in robotics technology.

With Disney’s involvement, theme park attractions can go beyond pre-programmed animatronics to feature robots capable of responding dynamically to guests. Meanwhile, Google DeepMind’s investment in robotics development shows that Newton could become a foundational tool for training and deploying robots across various industries.

Jensen Huang’s keynote did not stop at Newton. He also introduced Groot N1, a foundational model designed to improve humanoid robots’ ability to interpret and navigate their surroundings. Alongside this, Nvidia unveiled its upcoming AI chip roadmap, featuring the next-generation Blackwell Ultra and Rubin architectures, and introduced a new category of “personal AI computers.”

The models, Gemini Robotics and Gemini Robotics-ER, are based on Google’s Gemini 2.0 AI framework and are expected to enhance robots’ ability to perform tasks in industrial, commercial, and even domestic environments.

Gemini Robotics is a vision-language-action (VLA) model that allows robots to understand commands, interpret visual data, and execute physical actions. 

This means robots built with this model will be able to respond to spoken instructions, recognise objects, and interact with their surroundings more intuitively. 

The second model, Gemini Robotics-ER, takes this further by incorporating spatial reasoning abilities, enabling robots to understand complex environments and adjust their movements accordingly.

According to Google, these models have been designed to work with robots of various shapes and sizes, from humanoid robots to industrial machines commonly used in factories and warehouses. 

The company has already tested Gemini Robotics on its ALOHA 2 bi-arm robotic platform, demonstrating its ability to handle tasks requiring precision and adaptability. 

The model has also been successfully applied to Apptronik’s Apollo humanoid robot, which is being developed for real-world applications.

Google’s focus on robotics follows the recent decision by Figure AI to end its collaboration with OpenAI after making its own innovations in AI-powered robots. With the introduction of Gemini Robotics and Gemini Robotics-ER, Google is taking hold of the robotics industry.

One of the advantages of these models is their ability to reduce development costs for robotics startups. Companies can speed up the time it takes to bring functional robots to market when they leverage Google’s AI-powered frameworks. 

The tech giant has also stressed the flexibility of its models, stating that developers can customise them for specific robotic applications using Gemini’s advanced reasoning capabilities.

Google has partnered with Apptronik to integrate its AI models into humanoid robots, aiming to create machines capable of performing tasks that require both cognitive reasoning and physical dexterity. 

In August, Apptronik raised $350 million in a funding round led by B Capital and Capital Factory, with Google also participating to support the development of next-generation humanoid robots.

This is not Google’s first move in robotics. The tech giant previously acquired Boston Dynamics in 2013, a company famous for its dog-like and humanoid robots, before selling it to SoftBank in 2017. 

However, the launch of these new AI models is a renewed push into robotics, with Google focusing on AI-driven intelligence rather than hardware development.

With the integration of multimodal AI into robotics, Google aims to bridge the gap between digital intelligence and real-world functionality.

Hence, the increasing demand for automation in manufacturing, logistics, and service industries, will ensure Gemini-powered robots make work even more seamless.

This means users will have to pay $30 for a minute-long clip or $1,800 for an hour of generated video.

The cost was quietly listed on Google’s official pricing page, drawing comparisons to high-budget Hollywood productions. 

Jon Barron, a researcher at Google DeepMind, noted that the production budget for “Avengers: Endgame” was around $356 million—roughly $32,000 per second of footage. 

While Veo 2 is unlikely to produce feature-length blockbusters anytime soon, it is designed to create clips that extend beyond two minutes.

This pricing puts Veo 2 in competition with other high-end video generation models. OpenAI, for instance, offers access to its Sora model through a $200 monthly ChatGPT Pro subscription. 

However, OpenAI has not yet disclosed per-second pricing for Sora’s output, making direct cost comparisons difficult.

The market is growing for AI-generated video content and Google’s Veo 2 price at this level shows it can’t be left out. With increasing demand for high-quality synthetic media, businesses, advertisers, and content creators are expected to weigh the cost-effectiveness of such models against traditional video production methods.

Veo 2 is part of Google’s expansion into generative technology, which also includes its Gemini AI models. These models follow a tiered pricing system based on usage, ranging from text-based tasks to image and video generation. 

With AI-powered video creation climbing up innovatively, pricing will determine its accessibility and adoption across industries.