THE SHORT VERSION
- Google announced Gemini 4 Carbon on October 29, saying it ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59.谷歌 10 月 29 日发布 Gemini 4 Carbon，称其在 Artificial Analysis 智能指数 5.0 版中以 59 分排名第一。
- Google reported 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym and 71.0% on CWE-bench, and said the model runs long agentic coding tasks.谷歌公布 SEC-bench Pro 漏洞发现 88.2%、ExploitGym 54.8%、CWE-bench 补丁 71.0%，并称模型可执行长时间编程任务。
- Carbon is being released first to a subset of cyber defenders through Google's Fairwind Program, at the same introductory price as Argon, before wider access.Carbon 先通过 Fairwind 计划开放给部分网络安全防御者，定价与 Argon 相同，之后才扩大范围。
A frontier model pitched at defenders first面向防御方的前沿模型
Google announced Gemini 4 Carbon on October 29, describing it as its strongest model for cybersecurity and long-horizon work. The company's blog post said Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59, and that it builds on Gemini 4 Argon, which the post said launched two weeks earlier. Qatar's state news agency reported the launch the same day and placed Argon in early October, a small difference in dates that shows how quickly the company is now shipping frontier models.谷歌 10 月 29 日发布 Gemini 4 Carbon，称这是其在网络安全与长周期任务上最强的模型。公司博客称，Carbon 在 Artificial Analysis 智能指数 5.0 版中以 59 分排名第一，并建立在 Gemini 4 Argon 之上，博客称 Argon 于两周前发布。卡塔尔国家通讯社同日报道了这次发布，并称 Argon 于 10 月初推出。日期的细微差别说明，这家公司发布前沿模型的节奏已经很快。
The framing matters more than the headline number. Google is not presenting Carbon as a general chatbot upgrade first. It is presenting it as a tool for finding, validating and fixing vulnerabilities, and it is giving it to a narrow group before anyone else. Defenders first, developers and consumers later: that order is the substance of this launch, and it is why the benchmark scores are worth reading carefully rather than repeating.关键的是定位，而不是标题里的排名。谷歌没有先把 Carbon 当作通用聊天机器人升级来卖，而是把它当作发现、验证和修复漏洞的工具，并先交给一小批人。先给防御方，再给开发者与普通用户：这个顺序才是本次发布的实质，也是为什么这些基准分数值得细读，而不是照抄。
The security scores measure three different jobs三个安全分数衡量的是三件不同的事
Google reported 88.2% on SEC-bench Pro, which it describes as a test of finding real vulnerabilities in software; 54.8% on ExploitGym, which tests whether a model can turn a finding into a working exploit; and 71.0% on Collinear's CWE-bench, which tests patching. The company said the models were evaluated without guardrails, and QNA independently reported that Carbon topped both SEC-bench Pro and ExploitGym.谷歌公布，Carbon 在 SEC-bench Pro 上取得 88.2%，该测试衡量在软件中发现真实漏洞的能力；在 ExploitGym 上取得 54.8%，该测试衡量模型能否把发现转成可用的利用方式；在 Collinear 的 CWE-bench 上取得 71.0%，该测试衡量打补丁。公司称这些评估是在不启用护栏的情况下进行的，卡塔尔通讯社也独立报道 Carbon 在 SEC-bench Pro 与 ExploitGym 上居首。
Those are three stages of the same defensive job, and a model can be strong at one and weak at another. Finding a flaw, deciding whether it can actually be exploited, and writing a patch that does not break the system are different skills. An 88.2% figure is a rate on one benchmark, not a measure of how many real flaws Carbon would find in a given codebase, and the blog post does not claim otherwise. The useful question for a defender is which of the three stages costs their team the most time.这三项是同一项防御工作的三个阶段，模型可能擅长其中一个而弱于另一个。发现漏洞、判断它是否真的可被利用、写出不会破坏系统的补丁，是三种不同能力。88.2% 是一个基准上的比率，不等于 Carbon 在某个代码库中能找到多少真实漏洞，博客也没有这样声称。对防御者更有用的问题是：这三个阶段里，哪一个最耗团队的时间。
Long-horizon coding is the other half of the pitch长周期编程是这套说法的另一半
Carbon is also aimed at sustained engineering work rather than single answers. Google said its own teams have seen the model run for days while working through dead ends on multi-step tasks such as finding and patching subtle bugs across large codebases and low-level systems. On WebDev Arena, a public comparison of web development output, the company said Carbon gained 89 elo points to reach fourth place at 1768.Carbon 也瞄准持续性的工程任务，而不是一次问答。谷歌称其内部团队见过该模型连续运行数天，在多次走入死路后仍继续推进多步骤任务，例如在大型代码库和底层系统中定位并修复细微缺陷。在比较网页开发输出的公开榜单 WebDev Arena 上，公司称 Carbon 提升 89 个 elo 分，以 1768 分升至第四。
The stronger claims in this section are internal and unpublished. Google said Carbon delivers more than twice the success rate of Argon on generating secure code in its own benchmarks, that its Chrome team found Carbon generated high-quality patches more than ten times better than another lab's large frontier model, and that the cloud security firm Wiz sees Carbon leading on web penetration testing benchmarks. None of those comparisons can be checked from outside, and the models they are measured against are not named.这一节里更强的说法都是内部数据、未公开发表。谷歌称在其内部基准上，Carbon 生成安全代码的成功率是 Argon 的两倍以上；其 Chrome 团队认为 Carbon 生成高质量补丁的表现，比另一家实验室的大型前沿模型好十倍以上；云安全公司 Wiz 也认为 Carbon 在网页渗透测试基准上领先。这些比较无法从外部核实，被比较的模型也没有点名。
Access is narrow, and the price matches Argon开放范围很窄，价格与 Argon 相同
Google said it is taking part in the US government's voluntary process for pre-release model testing, and that from this week Carbon would go to a subset of cyber defenders, with access widened gradually to additional trusted partners through its Fairwind Program. QNA reported that the model is currently available to a limited group of cybersecurity companies, with wider availability to follow. Developers, businesses and consumers are described as coming later, once Google has gathered feedback and adjusted guardrails.谷歌称其参与了美国政府的前沿模型发布前自愿测试流程，并称从本周起先把 Carbon 交给一部分网络安全防御者，随后通过 Fairwind 计划逐步扩大到更多受信任的合作伙伴。卡塔尔通讯社报道，该模型目前只提供给有限的一批网络安全公司，之后才会扩大范围。开发者、企业与普通用户被列为更晚的阶段，要等谷歌收集反馈并调整护栏之后。
The commercial terms are unusually plain for a frontier release. Carbon will launch at the same introductory price as Argon: $2 per million input tokens, $10 per million output tokens, and cached input tokens at 95% off the input price. For a security team weighing whether to route vulnerability triage through a model, that is the number that decides whether the benchmarks turn into a budget line. Related coverage on this site includes Gemini 4 Argon's app rollout and Google Cloud's agent work.对一次前沿模型发布来说，商业条款少见的直白。Carbon 将以与 Argon 相同的首发价格推出：每百万输入 token 2 美元、每百万输出 token 10 美元，缓存输入 token 按输入价的 95% 折扣计价。对一个正在考虑是否把漏洞分流交给模型的安全团队而言，这个数字决定了基准分数能否变成预算条目。本站相关报道包括 Gemini 4 Argon 的应用开放与谷歌云的智能体工作。
Safety numbers ship with the model安全数字与模型一同发布
Google paired the capability claims with alignment numbers. It said Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, with attempts to cheat dropping to 0.4% while overall accuracy improved; that high-severity misalignment alerts in its internal live coding environments fell by almost 80%; and that unsupported claims were cut by roughly half compared with Argon.谷歌在能力说法之外，也给出了对齐方面的数字。它称 Carbon 把 Terminal-Bench 4.0 上的成功作弊率降到 0.0%，尝试作弊的比例降到 0.4%，同时整体准确率还有提升；在其内部实时编程环境中，高严重度失准告警下降近 80%；与 Argon 相比，无依据的说法减少约一半。
The post also reported that Carbon reduced overreach by 26% and under-service by 29%, and that it resisted every evaluated attempt on Gray Swan's indirect prompt injection benchmark, a 0.0% attack success rate against Argon's 0.7%. These are vendor-reported figures from internal and third-party evaluations, and they describe the model Google tested, not every deployment of it. The prior model's public safety reporting is the comparison point, and this site covered Anthropic's three-tier cyber verification plan as another approach to gating access.博客还称，Carbon 把越权行为减少 26%、服务不足减少 29%，并在 Gray Swan 的间接提示注入基准上抵御了所有被评估的攻击，攻击成功率为 0.0%，而 Argon 为 0.7%。这些都是厂商报告的数字，来自内部与第三方评估，描述的是谷歌测试的那个模型，而不是它的每一次部署。上一代模型的公开安全报告是对照点，本站也曾报道Anthropic 的三档网络安全验证计划，那是另一种限制开放范围的思路。
What is established, and what is still open哪些已经确定，哪些仍待回答
Established: Google announced Gemini 4 Carbon on October 29; it says the model leads the Artificial Analysis Intelligence Index at 59; it published security, coding and alignment scores; access starts with a small group of cyber defenders; and the introductory price matches Argon. Not established: whether outside evaluators will reproduce the security scores, how often the model finds real, previously unknown flaws in production code, and when the wider rollout actually happens.已经确定的：谷歌 10 月 29 日发布了 Gemini 4 Carbon；它称该模型以 59 分领跑 Artificial Analysis 智能指数；它公布了安全、编程与对齐方面的分数；开放从一小批网络安全防御者开始；首发价格与 Argon 相同。尚未确定的：外部评估者是否会重现这些安全分数、模型在生产代码中发现真实且此前未知漏洞的频率有多高，以及更大范围开放究竟何时发生。
For readers, three things are worth watching. The first is independent replication of the vulnerability-finding numbers on code the model has not seen. The second is whether the Fairwind cohort publishes what Carbon found and what it missed, which is more useful than another benchmark table. The third is the date access widens, because that decides whether this is a specialist tool or the next default model. Until then, the launch is best read as a claim with unusually specific numbers attached, aimed at a group that can test it.对读者而言，有三件事值得关注。第一是漏洞发现数字能否在模型没见过的代码上被独立重现。第二是 Fairwind 这批使用者是否公开 Carbon 找到了什么、漏掉了什么，这比再一张基准表更有用。第三是开放范围扩大的日期，因为它决定这是一件专业工具，还是下一个默认模型。在那之前，这次发布最好被读作一份附有异常具体数字的说法，交给一批有能力测试它的人来检验。
Sources & reporting limits来源与核查范围
Follow Industry to find related stories on your next visit.关注产业动态，下次访问时阅读相关文章。
