刚刚，Gemini 4 Carbon 正式发布，哈基米重回世界第一

刚刚，Gemini 4 Carbon 正式发布，哈基米重回世界第一

Gemini 终于又支棱起来了？

当全世界的目光都聚焦在苹果首款触控屏 MacBook Pro 身上，Google DeepMind 刚刚正式宣布推出新一代旗舰模型 Gemini 4 Carbon。

按照官方公布的数据，Gemini 4 Carbon 在最新 Artificial Analysis Intelligence Index v5.0 榜单中拿下 59 分，超越 Claude Opus 5.5 和 GPT-6 Astra，登上全球大模型榜首。

更惊人的是，Gemini 4 Carbon 已经能连续数天处理复杂工程任务，自主寻找、验证并修复软件安全漏洞，在多项网络安全基准测试中刷新纪录。

然而，如此强大的模型，普通用户暂时还用不上：谷歌决定优先向少数网络安全防御人员开放，其他用户何时能够体验仍没有明确时间表。

重回世界第一，Gemini 也开始卷能力上限了

回顾今年的大模型竞争，谷歌的更新节奏明显正在「知耻而后勇」。

此前，接连推出 Gemini Flash 3.5 至 3.8 系列后，谷歌迟迟没有兑现新一代旗舰 Pro 模型的期待。另一边，Anthropic 和 OpenAI 持续刷新能力上限，国产模型也在不断追赶，谷歌急需一款有分量的产品重新证明自己的实力。

直到两周前，Gemini 4 Argon 才终于登场。

Argon 意为「氩」，是一种性质稳定的惰性气体，常用于焊接和半导体制造中的保护环境。Argon 在文笔写作、表达质量等方面取得了明显提升，但距离全面领先顶级模型仍有差距。

如今，谷歌又带来了 Carbon，也就是「碳」。作为构成生命与有机化合物的重要元素，碳既能形成柔软的石墨，也能构成坚硬的金刚石。

真正让 Gemini 4 Carbon 引起关注的，还是它公布的一系列成绩。

就在 10 月 27 日，独立模型评测机构 Artificial Analysis 刚刚升级了 Intelligence Index v5.0。这套综合指数涵盖知识工作、编程、科学和通用推理四大领域，各占 25%，其中 60% 的评测权重采用私有测试数据或隐藏答案，进一步提高了测试难度。

昨天新版榜单发布时，Claude Opus 5.5 以 58 分排名第一，GPT-6 Astra 以 57 分紧随其后，GPT-6.1 Sol 则以 55 分位居第三。而 Gemini 4 Carbon 直接拿到了 59 分，比 Argon 高出 9 分，也以 1 分优势超过此前榜首的 Claude Opus 5.5。

不止 Artificial Analysis，Arena.ai 最新公布的榜单也显示，Gemini 4 Carbon（High）在 Text Arena 登顶第一、Agent Arena 排名第二、Code Arena WebDev 排名第四。

向左滑动查看更多内容

伴随着新模型登顶，谷歌内部也开始为 Gemini 积极造势。此前从 Anthropic 跳槽至 Google DeepMind 的华人研究员姚顺宇，同样为 Gemini 系列宣传打 call。

除了综合排名，Gemini 4 Carbon 在实际工程任务上的表现也有明显变化。

过去，大模型可以在几分钟内写出一段程序，但面对大型软件工程，问题就复杂得多。一个隐藏在数百个文件中的故障，可能需要开发者反复检查依赖关系、定位异常、修改代码，再经过多轮测试才能解决。

对于 AI 而言，难点除了编程能力，还有能否在长时间执行任务的过程中保持目标一致，并在失败之后继续尝试。

谷歌表示，Gemini 4 Carbon 已经能够连续运行数天，处理大型代码库和底层系统中的复杂工程问题，即使遇到障碍，也能持续寻找其他方案，最终定位并修复一些隐蔽的软件缺陷。

网页开发方面，Gemini 4 Carbon 在 WebDev Arena 上的 Elo 评分提升了 89 分，达到 1769 分，排名第四。

相比过去随叫随到的编程助手，谷歌希望 Gemini 4 Carbon 能够成为真正承担复杂工程任务的自主 Agent。

AI 开始自主寻找、利用和修复漏洞

如果说长程编程体现了 Gemini 4 Carbon 的工程能力，那么网络安全，可能才是谷歌此次更新最值得关注的部分。

传统的软件安全工作需要经历多个环节，首先从大量代码中寻找漏洞，再判断漏洞是否真的可以利用，最后针对问题编写修复程序，并验证修改是否有效。

Gemini 4 Carbon 的能力提升，恰好覆盖了这条完整流程。

在 SEC-bench Pro 测试中，Gemini 4 Carbon 取得了 88.2% 的成绩，主要考察模型在真实软件中自主发现复杂漏洞的能力；

在 ExploitGym 测试中，Gemini 4 Carbon 达到 54.8%，重点考察能否实际利用漏洞来验证风险；

在 Collinear 的 CWE-bench 测试中，Gemini 4 Carbon 又取得 71.0%，能够针对发现的问题生成有效的代码修复方案。

按照谷歌披露的结果，Gemini 4 Carbon 在上述三项测试中均达到了领先水平，相关成绩是在未启用安全防护限制的条件下测得。

谷歌内部测试也给出了相似结果。

在安全代码生成方面，Gemini 4 Carbon 的成功率超过 Argon 的两倍；Chrome 团队发现，其生成高质量安全补丁的表现比另一家实验室的前沿模型高出十倍以上，但未披露对比对象。谷歌旗下云安全平台 Wiz 也观察到，Carbon 在网络应用渗透测试中取得领先成绩。

从发现漏洞、验证风险到完成修复，AI 正在逐渐掌握网络安全专家的一整套工作流程。但同样的能力既能用于修复系统，也可能被攻击者利用，因此，谷歌还重点提升了 Carbon 在自主执行任务时的可靠性。

在 Terminal-Bench 4.0 测试中，Carbon 成功作弊的比例降至 0.0%，尝试作弊的比例仅为 0.4%，同时任务准确率有所提升。在实际编程环境中，高严重程度的行为不一致警报减少近 80%，缺乏证据支持的断言减少约一半，越权行动和未充分完成任务的情况也分别下降 26% 和 29%。

另一项进展来自提示词注入防护。

攻击者可能在网页或文档中植入恶意指令，诱导 AI 偏离原本任务，而在 Gray Swan 的间接提示词注入测试中，Carbon 将攻击成功率从 Argon 的 0.7% 进一步降至 0.0%，抵御了此次测试覆盖的全部攻击尝试。

当然，测试中的零成功率并不意味着现实中的所有攻击都能被阻止。有 Hacker News 网友认为，减少 AI 在执行任务时投机取巧，可能是提升前沿模型实用性的重要进步之一。

毕竟，当 AI 开始自主处理复杂工程任务，人们除了关心它能否完成工作，还必须考虑它会不会越权操作、编造结果，甚至在遇到障碍时擅自行动。

AI 圈最时髦的营销，是无实物表演

不过，Gemini 4 Carbon 发布后，最引人关注的争议，反而和模型能力没有太大关系。

按照官方安排，Gemini 4 Carbon 本周将首先提供给部分网络安全防御人员，再通过 Fairwind Program 逐步扩大至可信合作伙伴。谷歌还在参与美国政府自愿开展的模型发布前测试，至于普通开发者和消费者何时能够使用，暂时没有时间表。

价格方面，Gemini 4 Carbon 将沿用 Argon 的首发定价，每百万输入 Token 收费 2 美元，输出 Token 收费 10 美元，缓存输入优惠 95%。推广期结束后，标准价格将调整为输入 4 美元、输出 20 美元。

值得一提的是，谷歌发布新模型的消息传到 Hacker News，评论区很快热闹起来。

不少网友直言，自己越来越反感谷歌提前宣布新模型、却迟迟不开放的做法；也有网友表示，已经懒得点开谷歌的模型发布链接，因为最后总会发现产品暂时用不上。

不知道从什么时候开始，AI 厂商们似乎越来越擅长一种奇特的发布方式。

原本等打磨好就能直接端上桌的旗舰模型，现在非要套上一层能力太强怕毁灭世界的安全叙事外壳，隔三差五发篇通稿围猎大伙的注意力，结果到头来却迟迟用不上。

这种「无实物表演」的发布方式属实是被 Anthropic 带到沟里去了。

归根结底，争抢全球第一 AI 头衔的竞争，也确实越来越像一场营销游戏。在一个模型几乎每隔几天就会刷新的时代，发布越来越容易，真正让用户相信它有多强，反而越来越难。

毕竟，Benchmark 可以告诉我们模型在考试中有多厉害，但能否进入真实的工作环境、创造持续的价值，还得等用户真正用上它之后才知道。

而对于 Gemini 来说，或许比再次宣布全球第一更重要的事情，是尽快让这个全球第一让更多人有机会用上。

本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

Just now, Gemini 4 Carbon, official publish, Hakimi returns to the top of the world.

Gemini, at last?

When the eyes of the world focused on MacBook Pro, the Apple's prime touch screen, Google DeepMind had just officially launched a new generation of flagship models, Gemini 4 Carbon.

According to officially published data, Gemini 4 Carbon scored 59 points in the latest Artificial Analysis Inteligence v. 5.0, surpassing Claude Opus 5.5 and GPT-6 Astra to the top of the global LLM list.

More strikingly, Gemini 4 Carbon has been able to handle complex engineering tasks for several days, independently seeking, validating and repairing software security gaps and updating records in multiple network security benchmarking tests.

However, for the time being, such a powerful model has not been used by ordinary users: Google has decided to give priority to a small number of cyber-security guards, and when other users will experience it without clear timelines.

Back to the world, Gemini started rolling up the capacity limit.

Looking back at this year's LLM competition, Google's new rhythm is clearly “dignity and courage”.

Earlier, after the successive roll-out of the Gemini Flash 3.5 to 3.8 series, Google was late in fulfilling the expectations of the new-generation flagship Pro model. On the other side, Anthropic and Open AI are continuously recharging their caps, and the national production model is constantly catching up, and Google is in dire need of a piece of good quality to prove its strength again.

It was only two weeks ago that Gemini 4 Argon finally arrived.

Argon, which means "thorium", is an inert gas of a stable nature, often used to protect the environment in welding and semiconductor manufacturing. Argon has achieved significant improvement in the quality of writing, expression, etc., but there are still gaps in the overall lead lead model.

Now Google brings in carbon, carbon. As an important element of life and organic compounds, carbon can form both soft graphite and hard diamonds.

What really made Gemini 4 carbon cause concern is a series of results that it published.

On October 27, the independent modelling agency Artificial Analysis just upgraded its Inteligence Index v. 5.0. This comprehensive set of indicators covers the four main areas of knowledge work, programming, science and generic reasoning, each accounting for 25 per cent, with 60 per cent of the assessment weights using private testing data or hidden answers, making testing more difficult.

Yesterday's new edition of publish, Claude Opus 5.5 ranked first in 58, GPT-6 Astra followed closely by 57, and GPT-6.1 Sol ranked third in 55. Gemini 4 Carbon received 59 points directly, 9 points higher than Argon, and one point more than Claude Opus 5.5.

More than only Artificial Analysis, Arena.ai's recently published list shows that Gemini 4 Carbon (High) is number one in Text Arena, Agent Arena is number two, and Code Arena WebDev is number four.

Slide left to see more content

And with the new model coming to the top, Google is beginning to be active for Gemini. Former Chinese researcher Yao Sun-woo from Anthropic to Google DeepMind, also called for the Gemini series.

In addition to the overall ranking, Gemini 4 Carbon's performance in the actual engineering tasks has also changed significantly.

In the past, LLM had been able to write a program in minutes, but the problem was much more complex in the face of large-scale software engineering. A failure hidden in hundreds of documents may require repeated examination of dependency, location anomalies, modification of codes by the developers before it can be resolved through multiple rounds of testing.

For AI, it is difficult to add programming capacity to the ability to maintain unity of purpose in the course of long missions and to continue to try after failure.

Google stated that Gemini 4 Carbon had been able to run for several days to address complex engineering problems in large code libraries and bottom systems, and that even if obstacles were encountered, other options could be sought on an ongoing basis to eventually locate and repair some hidden software deficiencies.

With regard to web development, Gemini 4 Carbon has increased the Elo rating on WebDev Arena by 89 points to 1769, the fourth.

Google hopes that Gemini 4 Carbon will become an autonomous person with a real complex engineering task, compared to the programming assistant who was called on.

AI to begin to identify, exploit and repair gaps on its own

If the long-range programming reflects Gemini 4 Carbon’s engineering capabilities, cybersecurity may be the most interesting part of Google’s update.

Traditional software security requires multiple processes, starting with a search for loopholes from a large number of codes, then judging whether they are actually available, and finally developing problem-specific fixes and verifying whether the modifications are effective.

Gemini 4 Carbon's increased capacity coincides with the complete process.

During the SEC-bench Pro test, Gemini 4 Carbon achieved 88.2 per cent of the results, mainly looking at the ability of models to identify complex gaps in real software;

In the Expluit Gym test, Gemini 4 Carbon reached 54.8 per cent, focusing on the practical use of loopholes to validate risk;

In the CWE-bench test in Collinear, Gemini 4 Carbon obtained 71.0 per cent to generate an effective code repair program for identified problems.

According to Google's findings, Gemini 4 Carbon had reached a lead in all three of the above-mentioned tests, and the results were measured under conditions where no safety restrictions had been activated.

Google's internal tests gave similar results.

In terms of security code generation, Gemini 4 Carbon ' s success rate is more than twice as high as Argon ' s; the Chrome team found that it produces high-quality safety patches more than ten times more than the front model of another laboratory, but did not disclose the object of the comparison. Wiz also observed that Carbon had made a lead in web application penetration testing.

From the detection of loopholes and the validation of risks to the completion of the repairs, AI is gradually mastering a complete workflow process for network security specialists. However, the same capabilities can be used both to repair the system and to be used by the attackers, and Google has also focused on improving the reliability of Carbon in carrying out his own mission.

In the Terminal-Bench 4.0 test, the percentage of successful cheating by Carbon fell to 1.0 per cent, while the rate of attempted cheating was only 0.4 per cent, while the mission accuracy rate increased. In the physical programming environment, the number of reports of high levels of behavioural incoherence decreased by nearly 80 per cent, the number of allegations without evidence decreased by about half, and the number of ultra vires actions and incomplete completions decreased by 26 per cent and 29 per cent, respectively.

Another development came from the injection of prompt protection.

The attackers may have incorporated malicious instructions in their web pages or files to induce AI to deviate from its original mission, while in the indirect prompt injection test in Gray Swan, Carbon further reduced the attack on success rate from 0.7 per cent in Argon to 1.0 per cent, resisting all attempts to attack covered by the test.

Of course, zero success rate in the test does not mean that all attacks in reality can be stopped. Hacker News users believe that reducing AI speculation in mission assignments may be one of the important advances in enhancing the utility of front-line models.

After all, when AI begins to take ownership of complex engineering tasks, one has to consider, in addition to being concerned about its ability to complete its work, whether it will operate ultra vires, fabricate the results, and even act on its own when it encounters obstacles.

AI, the most fashionable marketing, no physical performance.

However, after Gemini 4 Carbon Z94Z, the most interesting controversy was not so much related to modelling capacity.

According to official arrangements, Gemini 4 Carbon will be provided this week first to some of the network security guards and then gradually expanded to credible partners through Fairwind Programme. Google is also participating in the United States Government ' s voluntary pre-testing of model publish, and there is no timetable for when ordinary developers and consumers will be able to use it.

In terms of prices, Gemini 4 Carbon will follow Argon's initial pricing at $2 per million input token, $10 per output token and 95 per cent cache input. At the end of the extension period, the standard price will be adjusted to enter $4 and output $20.

It is worth mentioning that news of the new Google publish model came to Hacker News, and the comment area quickly became popular.

A number of Internet users have spoken out against Google’s early announcement of new models and its slow opening; others have suggested that Google’s model publish is already too lazy to open, as it will eventually be discovered that the product will not be used for the time being.

I don't know when the AI manufacturers seem to be getting good at a strange publish approach.

The flagship model, which would have been able to get to the table once it had been sharpened, would have been too powerful to be able to destroy the world’s safe-story husks, and would have been used in the end five times before a complete article had been written.

The publish method of "no physical performance" was actually taken to the ditch by Anthropic.

In the end, the competition for the world's first-ever AI title has become more and more like a marketing game. In an era when models are refreshed almost every few days, publish becomes easier and more and more difficult to truly convince users of how strong it is.

After all, Benchmark can tell us how powerful the model is in the test, but it has to wait for the user to actually use it to get into the real work environment and create a sustainable value.

And for Gemini, perhaps more important than re-declaring the first thing in the world, is to give the first person in the world a chance to use it as soon as possible.

Auto-fetched and translated by AI Gem Box. All rights belong to the original author.