谷歌最强 AI 模型 Gemini 4 Carbon 登顶全球智能指数榜首 “专职”做网络安全Google's Most Powerful AI Model Gemini 4 Carbon Tops Global Intelligence Index, "Dedicated" to Cybersecurity


AI free token/AI免费积分

IT之家 10 月 29 日消息,谷歌正式发布新一代前沿大模型 Gemini 4 Carbon,主打业界顶尖的网络安全能力与长程任务处理能力。

谷歌最强 AI 模型 Gemini 4 Carbon 登顶全球智能指数榜首 “专职”做网络安全

谷歌称其为当下最强的模型,重点强化了网络安全能力和处理复杂、长周期任务的能力。该模型在最新版 Artificial Analysis Intelligence Index v5.0 中排名第一,综合得分 59。

Gemini 4 Carbon 首发入门定价与两周前发布的 Gemini 4 Argon 保持一致,缓存输入 Tokens 可享输入价 95% 折扣。

入门优惠期结束后,将恢复常规定价:

在长程编程任务上,Carbon 专为高难度、需要持续运行多日的智能体工作流外观,可以自主遍历死胡同、完成大型代码库与底层平台的根因定位、漏洞补丁编写等复杂多步骤工程任务。

在 WebDev Arena 前端全栈开发榜单上,Carbon 较上代提高 89 个 Elo 分,以 1,769 分位列第 4。

谷歌最强 AI 模型 Gemini 4 Carbon 登顶全球智能指数榜首 “专职”做网络安全

网络安全能力是本次升级的核心亮点,Carbon 在防御全生命周期的漏洞发现、验证、修复环节均达到业界领先水平:

谷歌最强 AI 模型 Gemini 4 Carbon 登顶全球智能指数榜首 “专职”做网络安全

漏洞发现:在 SEC-bench Pro 基准上达成 88.2% 的成绩,可自主挖掘真实软件中隐藏的复杂漏洞。

谷歌最强 AI 模型 Gemini 4 Carbon 登顶全球智能指数榜首 “专职”做网络安全

漏洞利用验证:在 ExploitGym 基准上取得 54.8% 的成绩,可自动判断漏洞是否可被利用,协助安全人员优先处理高风险漏洞。

谷歌最强 AI 模型 Gemini 4 Carbon 登顶全球智能指数榜首 “专职”做网络安全

根因补丁修复:在 CWE-bench 基准上取得 71.0% 的成绩,可直接将验证后的漏洞转化为可运行的修复代码。

内部测试数据表明,Carbon 生成安全代码的成功率较 Argon 提高超过 2 倍;Chrome 团队测试表明,其生成高质量补丁的表现比其他实验室的大前沿模型好 10 倍以上。

谷歌最强 AI 模型 Gemini 4 Carbon 登顶全球智能指数榜首 “专职”做网络安全

Gemini 4 Carbon模型示意图

安全对齐维度,Carbon 在 Argon 的多层安全体系基础上进一步强化:

发布节奏上,谷歌表示正配合美国政府的预发布模型自愿测试流程,从本周起先向一小部分网络安全防御人员开放 Carbon,后续将借助 Fairwind 计划逐步扩大可信合作伙伴的访问范围,在收集反馈、迭代完善防护机制后,再面向开发者、企业与普通消费者开放。

本文来源:https://www.ithome.com/1/019/240.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

IT Home October 29 news: Google has officially released its next-generation frontier large model, Gemini 4 Carbon, featuring industry-leading cybersecurity capabilities and long-horizon task processing capabilities.

Google's Most Powerful AI Model Gemini 4 Carbon Tops Global Intelligence Index,

Google calls it the strongest model currently available, with key enhancements in cybersecurity capabilities and the ability to handle complex, long-cycle tasks. The model ranks first in the latest Artificial Analysis Intelligence Index v5.0, with an overall score of 59.

Gemini 4 Carbon's introductory pricing matches that of Gemini 4 Argon released two weeks ago, with cached input tokens enjoying a 95% discount on input pricing.

After the introductory offer period ends, it will revert to standard pricing:

For long-horizon coding tasks, Carbon is designed for high-difficulty intelligent agent workflows that require continuous multi-day operation, capable of autonomously navigating dead ends, completing root cause analysis in large codebases and underlying platforms, writing vulnerability patches, and other complex multi-step engineering tasks.

On the WebDev Arena frontend full-stack development leaderboard, Carbon improved by 89 Elo points over the previous generation, ranking 4th with 1,769 points.

Google's Most Powerful AI Model Gemini 4 Carbon Tops Global Intelligence Index,

Cybersecurity capabilities are the core highlight of this upgrade. Carbon achieves industry-leading levels in vulnerability discovery, verification, and remediation across the full defense lifecycle:

Google's Most Powerful AI Model Gemini 4 Carbon Tops Global Intelligence Index,

Google's Most Powerful AI Model Gemini 4 Carbon Tops Global Intelligence Index,

Google's Most Powerful AI Model Gemini 4 Carbon Tops Global Intelligence Index,

Internal test data shows that Carbon's success rate in generating secure code has improved by more than 2 times compared to Argon; Chrome team tests show that its performance in generating high-quality patches is more than 10 times better than other labs' large frontier models.

Google's Most Powerful AI Model Gemini 4 Carbon Tops Global Intelligence Index,

Gemini 4 Carbon模型示意图

On the safety alignment dimension, Carbon further strengthens the multi-layered safety system built on Argon:

Regarding the release schedule, Google stated it is cooperating with the U.S. government's voluntary pre-release model testing process, opening Carbon to a small group of cybersecurity defenders starting this week. It will subsequently expand access to trusted partners through the Fairwind program, and after collecting feedback and iteratively refining protection mechanisms, it will be made available to developers, enterprises, and general consumers.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry