Google DeepMind发布Gemini 4 Carbon主打安全编码
Google DeepMind Introduces Gemini 4 Carbon, Its Top Model for Cyber Defense and Coding
全新旗舰级安全与编码模型发布,性能数据详实且定位明确,值得安全与开发领域从业者关注。
Google DeepMind has introduced Gemini 4 Carbon, its strongest model to date. Gemini 4 Carbon is a frontier model from Google built for long-horizon agentic coding and cybersecurity defense. It ranks #1 on the updated v5.0 Artificial Analysis Intelligence Index with a score of 59. For developers, the catch is access. Carbon starts with a vetted group of cyber defenders, not the public API.
Google DeepMind 推出了 Gemini 4 Carbon,这是其迄今为止最强大的模型。Gemini 4 Carbon 是 Google 构建的前沿模型,专为长周期智能体编码和网络安全防御而设计。它在更新后的 v5.0 Artificial Analysis Intelligence Index 中排名第一,得分为 59。对于开发者而言,关键在于访问权限。Carbon 首先面向经过审核的网络安全防御者群体开放,而非公共 API。
TL;DR
TL;DR(太长不看)
- Size: Parameter count and context window not disclosed. Its predecessor, Gemini 4 Argon, has a 1,048,576-token context.
- Runs on: Google-hosted only. No open weights, no self-hosting, no quantized variants.
- Performance: Leads 15 of 17 published benchmarks versus Argon, GPT-6 Astra, Opus 5.5 and Fable 5.1.
- Best: 96.2% BFS F1 on GraphWalks (256K to 1M tokens), versus 84.2% for Argon.
- Worst: 81.5% on OSWorld-2.1, trailing Claude Opus 5.5 at 87.2%.
- Bottom line: Best: state-of-the-art vulnerability finding, exploitation and patching. Worst: you probably cannot call it yet.
- 规模:参数量上下文窗口未公开。其前身 Gemini 4 Argon 拥有 1,048,576 token 的上下文窗口。
- 运行环境:仅限 Google 托管。无开源权重,不支持自托管,无量化变体。
- 性能:在 17 项已发布的基准测试中,有 15 项领先于 Argon、GPT-6 Astra、Opus 5.5 和 Fable 5.1。
- 最佳表现:在 GraphWalks(256K 至 1M tokens)上 BFS F1 得分达 96.2%,而 Argon 为 84.2%。
- 最差表现:在 OSWorld-2.1 上得分为 81.5%,落后于 Claude Opus 5.5 的 87.2%。
- 总结:最佳:最先进的漏洞发现、利用和修补能力。最差:你可能还无法调用它。
What is Gemini 4 Carbon?
什么是 Gemini 4 Carbon?
Carbon is the second Gemini 4 model, following Gemini 4 Argon. Google team states that the model targets hard, long-running engineering work. Internal teams saw Carbon run for days on single tasks. Examples include root-causing and patching subtle bugs across large codebases and low-level systems.
Carbon 是继 Gemini 4 Argon 之后的第二款 Gemini 4 模型。Google 团队表示,该模型旨在应对困难且耗时的工程工作。内部团队观察到 Carbon 在单个任务上连续运行数天。示例包括在大型代码库和低级系统中定位根本原因并修补细微 bug。
Google also tuned Carbon for full-stack web development and multimodal understanding. On WebDev Arena, Carbon gained 89 Elo points to reach #4 at 1769.
Google 还对 Carbon 进行了全栈 Web 开发和多模态理解的优化。在 WebDev Arena 上,Carbon 获得了 89 个 Elo 分,以 1769 分升至第 4 名。
How does Gemini 4 Carbon perform on benchmarks?
Gemini 4 Carbon 在基准测试中的表现如何?
Google published a 17-row comparison in its launch post. Carbon was run at the highest thinking setting. Most rival numbers are self-reported by each provider, per Google’s evaluation methodology.
Google 在其发布文章中公布了一项包含 17 行的对比数据。Carbon 是在最高思考设置下运行的。根据 Google 的评估方法,大多数竞争对手的数据由各提供商自行报告。
Agentic coding: Carbon scores 65.2% on Terminal-Bench 4.0 (AA), ahead of Opus 5.5 at 59.6%. It hits 9.6% on SWE-sweep, narrowly ahead of GPT-6 Astra at 9.5%. On FrontierSWE, GPT-6 Astra leads at 65.5% versus Carbon’s 61.2%.
智能体编码:Carbon 在 Terminal-Bench 4.0 (AA) 上得分为 65.2%,领先于 Opus 5.5 的 59.6%。在 SWE-sweep 上得分为 9.6%,略高于 GPT-6 Astra 的 9.5%。在 FrontierSWE 上,GPT-6 Astra 以 65.5% 领先,Carbon 为 61.2%。
Computer use and knowledge work: Carbon reaches 77.4% on Agents’ Last Exam v1.1. On Vending-Bench 2, it ends with a $16,484 bank balance, the highest listed.
计算机使用与知识工作:Carbon 在 Agents’ Last Exam v1.1 上达到 77.4%。在 Vending-Bench 2 上,其最终银行余额为 16,484 美元,为所列最高值。
Long context and science: GraphWalks is the widest gap: 96.2% versus 71.8% for GPT-6 Astra. Carbon also posts 80.0% on RiemannBench.
长上下文与科学:GraphWalks 差距最大:Carbon 为 96.2%,而 GPT-6 Astra 为 71.8%。Carbon 在 RiemannBench 上也取得了 80.0% 的成绩。
Why does Carbon relevant for cybersecurity?
为什么 Carbon 对网络安全相关?
Google frames Carbon around the full defense lifecycle: find, validate, then fix. Cyber benchmarks were evaluated without guardrails.
Google 将 Carbon 围绕完整的防御生命周期进行框架设计:发现、验证,然后修复。网络安全基准测试是在没有护栏的情况下进行评估的。
- Finding: 88.2% on SEC-bench Pro, versus 85.4% for GPT-6 Astra.
- Exploiting: 54.8% on ExploitGym, versus 42.4% for GPT-6 Astra and 34.5% for Opus 5.5.
- Patching: 71.0% on Collinear’s CWE-bench, versus 68.0% for Argon and GPT-6 Astra.
- 发现:在 SEC-bench Pro 上得分为 88.2%,而 GPT-6 Astra 为 85.4%。
- 利用:在 ExploitGym 上达到 54.8%,而 GPT-6 Astra 为 42.4%,Opus 5.5 为 34.5%。
- 修补:在 Collinear 的 CWE-bench 上达到 71.0%,而 Argon 和 GPT-6 Astra 均为 68.0%。
Google computed SEC-bench Pro and ExploitGym scores itself, using its Antigravity harness. ExploitGym ran with mitigations off. Internally, Carbon showed more than 2x higher success at generating secure code than Argon. Google team states that the Chrome team found Carbon over 10x better than another lab’s model at high-quality patches. Wiz reports Carbon leading its web penetration testing benchmarks.
Google 使用其 Antigravity 框架自行计算了 SEC-bench Pro 和 ExploitGym 的得分。ExploitGym 运行时未开启缓解措施。内部数据显示,Carbon 生成安全代码的成功率比 Argon 高出两倍以上。Google 团队表示,Chrome 团队发现 Carbon 在高质量补丁方面比另一家实验室的模型好 10 倍以上。Wiz 报告称 Carbon 在其 Web 渗透测试基准中领先。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力