Google DeepMind opened its strongest AI model to the public on 10 November 2026, and on the same day it showed a humanoid robot working on a live car factory line. The two launches landed together, and together they say something clear about where Google is placing its bets.
Carbon is built for the hardest jobs
Gemini 4 Carbon is the second frontier model in the Gemini 4 family, following Gemini 4 Argon. Google calls it the strongest model it has built, with frontier performance in cybersecurity and in long horizon agentic work across knowledge work, software engineering and scientific research.
“Today, we’re making Gemini 4 Carbon available to developers, enterprises, scientists, and consumers.”
Google DeepMind, in its Gemini 4 Carbon launch post
Argon stays the default for most software and knowledge work. Carbon is the model for work that tests the limits of intelligence, such as complex multi step coding and debugging and scientific research. It uses more tokens than Argon to explore thoroughly, reason deeply and run extended verification loops.
Google says Carbon ranks first on three independent workflow benchmarks, namely CursorBench 4.0, Perplexity WANDR and Zapier AutomationBench, at more than twice lower cost per task than the next best external model. Those rankings are Google’s own claims from its launch post, and independent checks have not yet been published.
What developers and businesses actually get
Paying developers and enterprises can use Carbon through the Gemini API in Google AI Studio and Gemini for business. Google AI Ultra subscribers get it in Google Antigravity, the Gemini app and AI Mode in Search. Scientists can apply for six months of free access through the Antigravity programme.
The launch price matches Argon at 2 dollars per million input tokens and 10 dollars per million output tokens. That price ends on 30 November 2026, and then rises to 4 dollars and 20 dollars. Free users get no access yet, with no date given for a wider rollout.
Google is also previewing a feature called swarm in Antigravity, which turns one prompt into a coordinated team of agents that can be tracked in real time. Developers can give Carbon the roles that need the deepest reasoning, such as architectural design and auditing, while lighter models handle parallel work.
The cyber case is the sharpest test
Carbon first reached a small group of cybersecurity defenders two weeks before the public launch, through Google’s Fairwind Program. In that time it was put to work on real security problems, and Google shared the results in its launch post.
Working with Wiz, Carbon scanned two major software providers and found previously unknown vulnerabilities in under two hours, work Google says would have taken weeks of manual research. The flaws were fixed with the affected companies in record time, limiting the possible impact on tens of thousands of companies.
Inside Google’s Project Fortify system, Carbon reached 100 percent recall on scanning and 95 percent precision on triage in internal testing. Its cloud research team used the model to prove that two bugs in open source software were exploitable, which led maintainers to reclassify them as critical and issue patches.
A robot that works a real factory line
Gemini Robotics 2.5 is Google DeepMind’s most capable vision language action model to date. It turns camera images and spoken instructions into robot movement, and Google says it learned new skills faster and handled more tasks than earlier models across more than 200 test tasks.
“Today, we are introducing Gemini Robotics 2.5, our most capable vision-language-action (VLA) model to date.”
Google DeepMind, in its Gemini Robotics 2.5 launch post
The headline test took place on active Mercedes-Benz production lines in Kecskemét, Hungary and Sindelfingen, Germany. Apptronik Apollo 2 humanoid robots sorted car parts into kits in the exact order each vehicle needed, reaching a 96 percent success rate. Google calls it a proof of concept, not a production deployment.
The model also brings new safety work. Google released an open benchmark called ASIMOV-Jailbreak, which tests whether robot AI can be tricked into unsafe actions. Its Gemini 4 Argon model resisted 99.5 percent of those attempts, against 82 percent for the earlier baseline.
Robotics 2.5 is rolling out to Google’s robotics partners only. There is no public API and no date for wider access, so most builders will watch this one from the sidelines for now.
Why this week matters
Google did not launch in a quiet week. Anthropic launched Claude Fable 5.2 on the same day, and OpenAI opened ChatGPT to business profiles the same day too. The three announcements together mark a week when the frontier labs all shipped at once.
Each lab is moving from a chatbot to a system that does work, whether that means agents in a coding tool, a business answering customers in ChatGPT, or a robot sorting parts on a factory line. Google is the only one whose launch spans screen and factory floor.
For startups, the practical read is simple. Carbon is available now for teams willing to pay, and its price doubles on 1 December. The robot side is still a partner programme, so the factory story is a signal of direction rather than a tool to build with today.
Google published the full details in its Gemini 4 Carbon announcement and its Gemini Robotics 2.5 announcement.
