product
2026-11-11 08:26
The release pace of large AI models has become so fast that people are overwhelmed.
Just now, Google DeepMind officially announced that its latest flagship model Gemini 4 Carbon is officially available to developers, enterprises, scientific researchers and ordinary consumers.
Carbon is currently the most comprehensive publicly available model in the Gemini family. It is particularly good at complex programming, network security, scientific reasoning, and multi-agent tasks that require long execution.
Google already made Carbon available to some cybersecurity defenders through its Fairwind program two weeks ago, and now it’s expanding its use even further.
Compared with the previously released Gemini 4 Argon, Carbon is mainly aimed at complex software engineering, cross-domain research and other difficult tasks, and can consume more tokens for in-depth reasoning, repeated verification and error correction, while Argon continues to undertake most daily programming and knowledge work.
Carbon ranks first in all three benchmarks: CursorBench 4.0, Perplexity WANDR, and Zapier AutomationBench, with a single-task cost more than half lower than the suboptimal performance of external competing models.
In terms of price, Carbon’s API preferential prices are the same as Argon’s, with USD 2 per million input tokens and USD 10 per million output tokens, which will be adjusted to USD 4 and USD 20 respectively from December 1.
Summon the AI team in one sentence to create a complete game overnight
An important upgrade of Gemini 4 Carbon is the ability to coordinate multiple AI Agents to jointly complete large-scale engineering projects.
To this end, Google has launched a preview version of the /swarm function in the AI programming platform Google Antigravity. Developers can enter a command to organize multiple AI Agents to perform tasks and view the progress in real time.
Among them, Carbon can be responsible for architecture design, task coordination and code review, while other faster and lower-cost models are responsible for specific development, and there is also a dedicated verification agent to check the modification results.
An open-world 3D game demonstrated by Google was completed by such an AI team.
Players assume the role of a pelican on a bicycle and freely explore Google's Mountain View campus. The development process spanned 3D modeling with Blender, game development with Three.js and WebGL, procedural sound production, and background music generation using Lyria.
Another demonstration was even more exaggerated.
Google provided Carbon with seven public city plan PDF files containing 206 engineering drawings, asking AI to rebuild the Google Platform 37 office building.
After 10 rounds of multi-agent tasks running overnight, Carbon finally produced an interactive Three.js 3D model containing approximately 906,000 triangles.
Carbon also created an interactive 3D historical exhibit based on historical photos and engineering drawings from the Eiffel Tower's construction period.
A clavichord with a 61-key mechanical structure was even restored based on CAD drawings and X-ray images.
In terms of web design, Carbon proposed five sets of visual design plans for the Google Merchandise Store through only two rounds of prompt words, and used Nano Banana 2.1 to generate product images.
Third-party developers also shared their actual experiences.
OpenCode software engineer Kit Langton said that the team had been unable to adapt the desktop application to the existing UI framework, so they submitted a list of requirements to Carbon before leaving get off work. When he came back the next day, Carbon had developed a Rust UI library from scratch and rewritten the OpenCode desktop application.
HubX asked Carbon to handle 16 real bugs in a code base with about 14,000 tests, and ultimately reported a 94% problem resolution rate and no new regression issues were introduced.
Zero-day vulnerability discovered in two hours, network security capabilities further improved
Cybersecurity is another core competency of Gemini 4 Carbon, with Carbon achieving leading results in vulnerability discovery, exploitation and remediation benchmarks such as SEC-bench Pro, ExploitGym and CWE-bench.
In Project Fortify, Google’s internal autonomous security system, Carbon’s vulnerability scanning recall rate reaches 100% and vulnerability classification accuracy reaches 95%. In the open source code warehouse test, its recall rate has achieved a relative improvement of more than 25% compared to Argon.
Even more striking is that in actual testing, Gemini 4 Carbon helped Wiz discover zero-day vulnerabilities at two large SaaS service providers in less than two hours, shortening the original research work from weeks to just a few hours.
The Google Cloud team used it to verify the exploitability of two open source software vulnerabilities, while SentinelOne used Carbon to discover four additional valid vulnerabilities that other reference models failed to identify.
As network security capabilities improve, Google has also strengthened model security protection.
In simulated network attack tests, Carbon successfully blocked 99% of attack paths launched by adaptive red team agents, and the proportion of single attacks that successfully breached protection dropped from 3.1% of Argon to 0.8%.
Cryptography verification completed in several weeks, new breakthrough in AI protein research
In terms of scientific research, Google announced three internal scientific research cases, involving cryptographic formal verification, biological experimental analysis and protein structure prediction.
The most noteworthy one is the code verification of the Signal encryption protocol.
Google researchers have developed an intelligent agent formal verification framework called AlphaProof Code, which uses models such as Carbon to convert security goals into mathematical theorems and then generate millions of lines of Lean proof code.
In the end, the team completed a mathematical proof of the key security properties of the code related to the Signal encryption protocol in a matter of weeks, while similar work in the past usually took a professional team several years.
In the field of biological experiments, Carbon can identify anomalies from experimental records and raw data, such as insufficient automatic mixing of deep well plates, residual trace sample contamination in the instrument, etc., helping researchers improve experimental reproducibility.
In terms of protein structure prediction, Google uses Carbon to form a multi-agent system to automatically find distantly homologous proteins in large-scale sequence and structure databases, providing more reference information for structure prediction models such as AlphaFold 3.
On seven low-homology protein families in the CASP17 blind test dataset, Carbon found an average of 40-fold increase in the number of homologous sequences. Relevant results are expected to help researchers improve the prediction capabilities of some difficult-to-research proteins without retraining the structure prediction model.
From writing code to doing financial analysis, AI begins to take on more complete tasks
In addition to technology research and development, Carbon is also expanding into professional fields such as finance, law and corporate operations.
Google collaborated with Artificial Analysis to launch the Multimodal Knowledge Agent (MKA) benchmark test, which specifically evaluates the ability of AI Agents to process PDFs, spreadsheets, charts, and engineering drawings.
The Carbon ranked second in this test, but the average cost per task was more than six times lower than the top-ranked model.
In financial scenarios, Carbon tied with the best-performing model in Hebbia's end-to-end financial agent evaluation and ranked first in Balyasny Asset Management's comprehensive financial intelligence test covering 23 cutting-edge model configurations.
Hebbia says Carbon can produce more clearly structured Excel financial models, as well as higher-quality PowerPoint presentations.
Enterprise search company Glean found that Carbon can find specified meeting minutes, emails or business data from office data and directly generate analysis results and action recommendations.
When can non-paying users use it?
With the official opening of Gemini 4 Carbon, Google also announced how to use it for different users.
Developers can call Carbon through the Gemini API paid tier of Google AI Studio and Android Studio, and enterprise users can use it through Gemini for business and related development tools.
For ordinary consumers, Gemini 4 Carbon is now available to Google AI Ultra subscribers, covering Gemini App and Google Search AI Mode.
Researchers can apply for Google's previously launched free research program, and eligible participants will receive six months of high-usage access to Google Antigravity, using both Carbon and Argon models. Cybersecurity researchers will need to apply for the dedicated Carbon for cyber through the Fairwind Program.
Looking back at the consumer electronics market in the past, mobile phone manufacturers usually have a relatively fixed release cycle for their flagship products. A few important launches every year are enough to keep technology enthusiasts discussing for months.
Today's AI industry is a different story. New models with different characteristics appear almost every day. Some refresh the programming list, some overcome mathematical problems, and some can work for dozens of hours continuously.
It’s hard to imagine that we have entered an era where there are too many models to experience.
Such intensive technical updates are certainly a good thing for users, but after reading the release announcement of Gemini 4 Carbon, I still feel a little unfinished. After all, Google has demonstrated so many eye-catching capabilities, but ordinary consumers need to subscribe to Google AI Ultra to experience it, which is somewhat uninteresting.
Shan Zong
Hello, World Hello, Word
A total of 509 articles have been published
This article comes from the column
Previous article2026-11-11 08:20
Morning Post | Huawei's small foldable computer MateBook Fold S debuts/"GTA VI" starts preloading, targeting 30fps on all platforms/Dong Mingzhu responds to Gree's declining performance: Copper prices are rising, Gree will continue to use copper
Aifaner App
Love style, make the future within reach
Digital trendy brands focusing on tomorrow’s products
Advanced tools, prophets first, AIGC’s inspiration guide
A new era of car manufacturing, tomorrow’s travelers
Explore new lifestyles in the city and be your guide to tomorrow’s life
Copyright © Guangzhou Liwo Zhiyuan Investment Management Partnership (Limited Partnership) 2008 - 2026. Use of Aifaner website content for commercial purposes requires permission. Non-commercial use is licensed under CC BY-NC 4.0.
All content is made available under the CC BY-NC 4.0 for non-commercial use. Commercial use of this content is prohibited without explicit permission.
