正文 · AI 翻译
Thanks to IT Home readers 星河之梦, 软媒用户1053564, Nextpage, 秦淮一梦, and 软媒用户1830083 for the tips!
IT Home reported on November 11 that Google announced on November 10 local time the official availability of its Gemini 4 Carbon model for developers, businesses, researchers, and consumers.
Koray Kavukcuoglu, Google senior vice president and DeepMind chief AI architect, said it is Google's most capable reasoning model to date, aimed primarily at complex software engineering, multistep agent tasks, scientific research, and cybersecurity.
Gemini 4 Carbon was first announced on October 28 and had previously been made available to some cybersecurity defenders through the Fairwind Program, a cybersecurity collaboration program. Its availability has now been expanded.
Google said that, compared with Gemini 4 Argon, which launched around the same time, Carbon can devote more reasoning tokens to complex tasks to explore hypotheses, verify results, and check edge cases, making it better suited to multistep coding, complex debugging, and open-ended scientific research.
For developer tools, Gemini 4 Carbon is available to paying developers through the Gemini API and Google AI Studio, as well as to enterprise customers. The model is also being rolled out gradually to Google AI Ultra subscribers across Google Antigravity, the Gemini app, and AI Mode in Google Search.
According to the official Google AI Studio documentation, Gemini 4 Carbon's model ID is gemini-4-carbon. It accepts text, image, video, audio, and PDF inputs and produces text output.
The model has a unified context window of 1,048,576 tokens. Standard synchronous requests allow a maximum output of 65,536 tokens, while Long Decode can extend the maximum output to 1,048,576 tokens. The model also supports 3 reasoning effort levels: low, medium, and high.
In terms of pricing, Gemini 4 Carbon and Gemini 4 Argon have the same introductory promotional rates: $2 per million input tokens and $10 per million output tokens. According to Google AI Studio documentation, the promotional rates run through November 30, 2026, after which prices will change to $4 per million input tokens and $20 per million output tokens.
On coding and agent capabilities, Google said Carbon can handle tasks such as migrations across multiple software packages, fault localization across subsystems, and complex architectural refactoring. It can also complete full-stack application development through cycles of building, running, verifying, and fixing. Google is also previewing the /swarm feature in Google Antigravity, allowing developers to coordinate multiple AI agents with a single prompt to take on work such as architecture design, task coordination, and code auditing, and to execute tasks in parallel in separate workspaces.
Cybersecurity is one of the main applications of Gemini 4 Carbon. Test results published by Google show that the model scored 88.2% on the SEC-bench Pro vulnerability discovery test, 54.8% on the ExploitGym vulnerability exploitation test, and 71.0% on the CWE-bench vulnerability repair test. Google said these capabilities can be used to identify software vulnerabilities, verify whether they can be exploited, and generate fixes that address their root causes.
Google also described Carbon's use in real-world security research. Its Cloud Vulnerability Research team used the model to verify that 2 defects in open-source software were exploitable vulnerabilities, helping maintainers reclassify the issues as serious vulnerabilities and subsequently release fixes. While scanning 2 major SaaS providers, the Wiz team used Carbon to discover zero-day vulnerabilities in less than 2 hours and worked with the affected companies to fix them.
In scientific research, Google said Carbon can handle complex reasoning tasks in physics, mathematics, and the life sciences. In protein structure prediction research, researchers used the model to search for distant protein homologs in large archives of unannotated sequences, catalogs of environmental and viral sequences, and 3D structure databases. Across tests on 7 low-homology protein families in the CASP17 blind test set, Carbon increased the average number of homologous sequences retrieved by 40 times, providing more information for subsequent structure prediction.
Regarding safety protections, Google said Carbon builds on Gemini 4 Argon's safety mechanisms with stronger defenses against prompt injection and misuse for cyberattacks. In Gray Swan's indirect prompt injection (IPI) benchmark, the model's attack success rate was 0%. However, Google also acknowledged that the model's safety defenses have room for further improvement against complex attacks in certain distributions.
Google positions Gemini 4 Carbon as a high-reasoning model for complex tasks that run over long periods, with a different role from Gemini 4 Argon, which focuses on most everyday software engineering and knowledge work scenarios.
References:
“Gemini 4 Carbon”
“What's new in Gemini 4 Carbon”
“Gemini 4 Carbon - Model Card — Google DeepMind”
“Gemini 4 Carbon is now available”
Related reading:
“Google's most powerful AI model: Gemini 4 Carbon tops the global intelligence index and ‘specializes’ in cybersecurity”
“Google's most powerful AI model: Gemini 4 Argon launches with a 1 million-token output limit, targeting the toughest problems”
Advertising disclosure: External links in this article (including but not limited to hyperlinks, QR codes, and access codes) are provided to convey more information and save readers time in evaluating options. Results are for reference only. All IT Home articles containing external links include this disclosure.