谷歌发布Gemini 4 Carbon：迄今最强前沿模型，网络安全三项基准居首，智力指数59分居首
谷歌（Google）10月28日在官方博客宣布推出新一代前沿模型 Gemini 4 Carbon，称其为公司迄今最强的模型，在网络安全与长时程任务能力上达到业界领先水平。Carbon 在最新发布的 Artificial Analysis Intelligence Index v5.0 中以59分位居第一。
网络安全方面，Carbon 在三项外部基准上均居首位：发现真实漏洞的 SEC-bench Pro 得分88.2%，验证漏洞可利用性的 ExploitGym 得分54.8%，从根源修复漏洞的 CWE-bench 得分71.0%。谷歌称，其内部测试中 Carbon 生成安全代码的成功率较上一代 Gemini 4 Argon 提升逾一倍，Chrome 团队报告其生成高质量补丁的表现比另一家实验室的大型前沿模型高出10倍以上。
谷歌表示，Carbon 承接两周前发布的 Gemini 4 Argon，面向更难的长时程工程任务设计，可连续运行数日处理大型代码库中的复杂问题；在 WebDev Arena 榜单上，Carbon 得分升至1768分、排名第四。在安全性上，谷歌称 Carbon 在 Terminal-Bench 4.0 中成功作弊率降至0.0%，在 Gray Swan 间接提示注入基准中攻击成功率为0.0%。
发布节奏上，谷歌采取审慎推进：本周起先通过 Fairwind Program 向一部分网络防御者开放，再逐步扩大到更多可信伙伴，并正参与美国政府的自愿发布前模型测试流程；之后将尽快向开发者、企业与消费者开放。定价与 Argon 的上市初期价格相同：输入每百万 token 2美元、输出每百万 token 10美元，缓存输入按输入价格的95%折扣计费。
Google Unveils Gemini 4 Carbon, Its Strongest Frontier Model, Topping the Intelligence Index at 59 With State-of-the-Art Cybersecurity Scores
Google announced on October 28 on its official blog the launch of Gemini 4 Carbon, a new frontier model it calls its strongest yet, with state-of-the-art cybersecurity and long-horizon capabilities. Carbon ranks No. 1 on the newly updated Artificial Analysis Intelligence Index v5.0, scoring 59.
On cybersecurity, Carbon leads three external benchmarks: 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for validating exploitation risk, and 71.0% on CWE-bench for patching flaws at the root cause. Google said Carbon more than doubled the success rate for generating secure code over Gemini 4 Argon in internal benchmarks, and that its Chrome team saw Carbon perform more than 10 times better than another lab's large frontier model at generating high-quality patches.
Building on Gemini 4 Argon, launched two weeks ago, Carbon is designed for harder, long-horizon engineering work and can run for days on complex problems across large codebases. It rose to No. 4 on WebDev Arena with 1,768 points, up 89. On safety, Google said Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0% and resisted every evaluated attack on the Gray Swan Indirect Prompt Injection benchmark, also 0.0%.
Google is taking a deliberate rollout approach: starting this week, Carbon will be available to a subset of cyber defenders through its Fairwind Program, expanding gradually to additional trusted partners, and Google is engaged in the U.S. government's voluntary process for pre-release model testing. Access for developers, businesses, and consumers will follow. Pricing matches Argon's introductory levels: $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off.
来源：GSM·美国（原文链接）
图片来源：图：谷歌 Google 标志（公共领域，经 Wikimedia Commons）。 / The Google logo (public domain, via Wikimedia Commons).
