谷歌又把大模型排行榜的第一名抢回来了。
距离 Gemini 4 Argon 面向更广泛用户开放仅过去约两周，谷歌便推出了新一代前沿模型 Gemini 4 Carbon。这一次，谷歌不仅将目标对准了大模型综合智能榜首，更把编程、长程智能体任务和网络安全能力推到了核心位置。
谷歌发布 Gemini 4 Carbon
10 月 28 日，谷歌 CEO 皮查伊也在 x 上宣布了该消息，他表示，继两周前推出 Gemini 4 Argon 后，Gemini 4 Carbon 再次刷新了模型能力上限，以 59 分位居 Artificial Analysis 最新 v5.0 智能指数榜首，并在漏洞发现、漏洞利用和补丁修复三项网络安全测试中取得领先成绩。
相比 Argon，Carbon 在长程任务执行能力和安全防护方面进一步提升，目前已开始向部分可信网络安全防御专家开放，后续将通过 Fairwind 计划逐步扩大访问范围，最终面向公众开放。
谷歌表示，Carbon 的设计初衷是应对难度更高、周期更长的工程难题，它专为持续的智能体工作流程而构建。在谷歌内部，团队曾亲眼见证 Carbon 连续运行数天，突破重重难关，最终解决复杂的、多步骤的工程任务，例如在大型代码库和底层系统中查找并修复细微的 bug。
这次发布的新模型专注于提升模型在多个领域的性能，包括编码、智能体任务、科学研究任务，以及全栈 Web 开发和多模态理解。
在 Artificial Analysis Intelligence Index v5.0 中，Gemini 4 Carbon 以 59 分登顶第一，超过 Claude Opus 5.5 的 58 分和 GPT-6 Astra 的 57 分。与上一代 Gemini 4 Argon 的 50 分相比，Carbon 提升了 9 分。
但更值得关注的是编程能力。Artificial Analysis 的独立评测显示，Gemini 4 Carbon 已经跃居 Coding Agent Index 第二名，仅落后 Claude Opus 5.5 三分。
在网络安全领域，谷歌公布的三组数字同样引人注目：漏洞发现测试达到 88.2%，漏洞利用验证达到 54.8%，漏洞修复测试达到 71.0%。谷歌称，Carbon 在这三个关键基准中均取得领先成绩。
更有意思的是，这款综合评分登顶、具备更强漏洞挖掘能力的模型，普通开发者暂时还用不上。
谷歌决定首先向少量网络安全防御专家开放 Carbon，再通过 Fairwind 计划逐步扩大测试范围。
一边是更激烈的模型性能竞争，一边是更加谨慎的开放策略。Gemini 4 Carbon 的发布，正在呈现前沿大模型竞争的另一个侧面：当 AI 越来越能够独立完成复杂工程任务，模型能力越强，如何控制其实际行为就越重要。
综合榜单登顶，编程能力距离 Claude Opus 5.5 只差三分
此次 Gemini 4 Carbon 最直接的看点，是在第三方综合评测中重新取得领先。
Artificial Analysis 近期更新了 Intelligence Index，将其升级至 v5.0。新版指数整合九项评测，涉及专业知识工作、复杂编程、科学推理、事实性知识和智能体自动化等任务，试图用更复杂的真实任务衡量模型能力，而不仅仅测试知识问答或者单轮推理。
根据 Artificial Analysis 的评测，Carbon 在九项测试中的四项取得第一，包括 HLE-Diamond、Terminal-Bench-Science、AA-Omniscience 和 AutomationBench-AA。
这些测试覆盖了高难度知识与推理、科学环境中的终端操作、事实性知识，以及跨工具自动化任务。
不过，综合榜单第一并不意味着 Carbon 在所有细分能力上都超过竞争对手。
在 Artificial Analysis Coding Agent Index 中，使用 Antigravity CLI 的 Gemini 4 Carbon 高推理配置获得 59 分，而 Claude Code 搭配 Claude Opus 5.5 最高推理配置获得 62 分。Carbon 排名第二，双方差距缩小至三分。
其中，在 FrontierCode v1.1 测试中，Carbon 获得 51%，Claude Opus 5.5 获得 56%。这说明 Anthropic 在部分真实软件工程任务中仍然保持优势。
另一方面，在衡量专业知识工作的 AA-Briefcase 基准中，Carbon 获得 1609 Elo，较 Argon 的 1483 Elo 提高 126 分。其中，演示文稿任务提升 293 Elo，分析质量提升 120 Elo。
但 Carbon 在这项测试中仍落后于部分竞争模型，包括 Grok 4.7、Muse Wonder 和 Kimi K3.1 Pro。
换句话说，谷歌已经取得综合智能指数第一，但在编程智能体、专业知识工作等细分领域，领先位置依然处于激烈争夺之中。
值得注意的是，谷歌在官方公告中还特别强调了 Carbon 的网页开发能力。
在 WebDev Arena 上，Carbon 的 Elo 较此前提升 89 分，达到约 1768 分，排名第四。谷歌将这一变化归因于模型生成的网页成果质量改善，以及全栈开发、多模态理解和复杂工程能力的提升。
从代码生成，到终端执行，再到完整的软件工程任务，Carbon 的改进方向相当明确：它不再只追求生成一段正确代码，而是要让模型能够在真实的开发环境中持续完成任务。
这种变化，也正在成为当前旗舰模型竞争的核心。
谷歌称 Carbon 能连续工作数天
过去两年，大模型的编程能力经历了明显变化。
最初，开发者关注的是模型能否写出一个函数、修复一个 Bug，随后逐渐转向让 AI 理解代码仓库、调用工具、执行测试，甚至独立交付完整的软件项目。
如今，前沿模型开始面临一个更加复杂的问题：如果一项工程任务需要数小时乃至数天才能完成，AI 能否在反复失败、调试和修改的过程中坚持下去？
这也是 Carbon 此次重点强化的方向。
谷歌在官方技术介绍中披露，在内部工程环境中，研究人员已经观察到 Carbon 持续运行数天的情况。它可以在大型代码库中追踪隐蔽错误，分析底层系统问题，并尝试形成可运行的修复方案。
这类任务通常并不存在一条明确的解决路径。开发人员可能需要阅读大量代码，复现问题，尝试不同修复方式，再根据测试结果调整判断。
对于智能体而言，真正的难点不仅在于生成代码，更在于能否长期保持对任务目标、执行状态和工程约束的理解。
谷歌表示，Carbon 专门针对这类长程任务进行了优化，强调模型在遇到失败路径后继续分析和寻找解决方案，并不是简单结束任务。
这种能力方向，很大程度上反映了 AI 编程工具的发展趋势。
在传统的代码补全场景中，开发者可以即时检查模型输出；而在长程智能体任务中，模型往往需要独立进行数十次甚至上百次工具调用。一旦早期某个判断出现偏差，后续执行就可能逐渐偏离原始目标。
因此，模型能否及时发现错误、在失败后重新规划、明确自己尚未完成的任务，开始变得与编程能力同样重要。从这一角度来看，Carbon 的技术重点是试图让 AI 在复杂工程中完成更长、更可靠的工作闭环。
如果说复杂编程是 Carbon 最受开发者关注的能力，那么网络安全很可能是谷歌此次发布中最具战略意义的方向。
谷歌公布了部分内部工程数据。
在安全代码生成任务中，Carbon 的成功率达到 Argon 的两倍以上。Chrome 团队的内部测试也显示，Carbon 在生成高质量补丁方面，相比另一家实验室的大型前沿模型取得了超过十倍的表现提升。
性能登顶，价格却没涨
在模型性能持续提升的同时，谷歌选择让 Carbon 沿用 Argon 的首发定价。
按照官方公告，Carbon 初期 API 价格为每百万输入 Token 2 美元、每百万输出 Token 10 美元，缓存输入价格享受 95% 折扣。
但这里有一个容易被忽略的细节：这只是首发优惠价格，并非永久定价。
谷歌在官方公告的脚注中明确指出，首发优惠结束后，Carbon 将恢复至每百万输入 Token 4 美元、每百万输出 Token 20 美元。谷歌尚未公布 Carbon 优惠期的具体结束日期。
单纯从价格来看，Carbon 相比 Argon 没有上调单位 Token 成本，但性能有明显提升。
不过，对于运行复杂智能体任务的开发者而言，Token 单价并不能完全反映实际成本。
模型能否一次性完成任务、是否需要反复修改、执行过程中消耗多少推理 Token，以及工具调用和运行时间，都会影响整体费用。
Artificial Analysis 的评测提供了一个更具体的视角。在其综合智能指数测试中，按照标准价格计算，Carbon 的平均单任务成本约为 7.76 美元，而 Claude Opus 5.5 最高推理配置约为 10.04 美元。
这意味着，在该特定测试条件下，Carbon 的平均单任务成本比 Claude Opus 5.5 低约 23%。
Carbon 能为谷歌赢回资本信任吗？
颇具戏剧性的是，就在谷歌发布 Gemini 4 Carbon、重新夺回全球大模型综合性能榜首的同一天，其母公司 Alphabet 却遭遇了资本市场的冷遇。
模型发布同日，Alphabet 公布了 2026 年第三季度财报。数据显示，公司当季营收达到 1275.8 亿美元，同比增长 25%，但低于华尔街约 1279.7 亿美元的预期；净利润为 261.7 亿美元，同比下降约 25%。财报公布后，Alphabet 股价一度在盘后交易中下跌超过 10%。
值得注意的是，谷歌的 AI 业务并非没有增长。第三季度，Google Cloud 营收达到 307.1 亿美元，同比增长 103%，创下新的增长纪录。谷歌 CEO Sundar Pichai 还透露，Gemini 应用用户规模已经突破 10 亿。
真正引发投资者担忧的，是持续膨胀的 AI 投资成本，以及这些投入何时能够转化为足够可观的利润。
据路透社报道，Alphabet 第三季度自由现金流为负 90.7 亿美元，较上一季度负 59 亿美元进一步恶化。公司同时预计，2027 年资本支出还将显著增加。与此同时，谷歌搜索业务收入增速从上一季度的 17% 放缓至 12%，广告业务的增长前景也受到关注。
当然，Alphabet 的利润下降并不能完全归咎于 AI 投资。其当季还确认了约 87 亿美元的股权投资损失，对净利润及每股收益产生明显影响。
但这一幕仍然揭示了当前 AI 行业面临的现实矛盾。
一方面，Gemini 4 Carbon 的发布证明谷歌仍有能力在前沿模型竞争中取得领先；另一方面，训练更强的模型、建设更庞大的算力基础设施、维持高速的技术迭代，也意味着越来越高的资金投入。
对谷歌而言，赢得大模型排行榜第一名，和赢得资本市场的信任，正在成为两场不同的比赛。
参考链接：
https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Google has taken back the top of LLM's list.
Only about two weeks after Gemini 4 Argon was opened to a wider audience, Google launched a new frontier model, Gemini 4 Carbon. This time, Google not only targeted the top of the LLM integrated intelligence list, but also pushed programming, long-range Z42Z missions and network security capabilities to the core.
Google publish Gemini 4 Carbon
On 28 October, the news was also announced on x by Google CEO Pichai, who stated that, following the launch of Gemini 4 Argon two weeks ago, Gemini 4 Carbon had again updated the model capacity ceiling at 59 to top the latest V5.0 smart index of Artificial Analysis and had led the way in three network security tests in which loopholes were discovered, loopholes used and patched.
Compared to Argon, Carbon, which has further enhanced its long-range mission implementation capacity and security protection, has now begun to be opened to some credible cybersecurity defence experts, following up with the Fairwind plan to gradually expand access to the public.
Google stated that Carbon was originally designed to deal with more difficult, more pro-cyclical engineering challenges and was designed specifically for the ongoing AI agentQ52 process. Within Google, the team has witnessed the continued operation of Carbon for several days, breaking through difficult times and ultimately resolving complex, multi-step engineering tasks, such as locating and repairing fine bugs in large code libraries and bottom systems.
This time, the new model of publish focuses on enhancing the performance of the model in a number of areas, including coding, AI agent missions, scientific research missions, and whole house Web development and multimodular understanding.
In Artificial Analysis Inteligence v. 5.0, Gemini 4 Carbon tops at 59, exceeding 58 for Claude Opus 5.5 and 57 for GPT-6 Astra. Compared to 50 points from the previous generation Gemini 4 Argon, Carbon increased by 9 points.
But even more important is programming capacity. An independent assessment by Artificial Analysis indicates that Gemini 4 Carbon has jumped to the second place of Coding Age Index, only after Claude Opus 5.5.
In the area of cybersecurity, the three sets of figures published by Google are equally striking: 88.2 per cent for detection of loopholes, 54.8 per cent for validation of loopholes and 71.0 per cent for repair. According to Google, Carbon is ahead of all three key benchmarks.
More interestingly, the model, with the combined rating peaking and greater capability to exploit loopholes, has not been used by ordinary developers for the time being.
Google decided first to open up to a small number of cyber-security and defence experts, and then to gradually expand the scope of testing through Fairwind.
On the one hand, there is more competition for model performance and on the other hand, a more cautious approach to openness. Gemini 4 Carbon ' s publish is on the other side of the front line of LLM competition: as AI becomes more and more able to perform complex engineering tasks on its own, the greater the modelling capacity, the more important it is to control its actual behaviour.
Claude Opus 5.5 short of three points.
Gemini 4 Carbon's most direct point of view is to regain the lead in a comprehensive third-party assessment.
Artificial Analysis recently updated Inteligence Index to V5.0. The new version combines nine assessments, involving tasks such as professional work, complex programming, scientific reasoning, factual knowledge and AI agent automation, and attempts to measure model capabilities with more complex real task, rather than simply testing knowledge questions and answers or single-cycle reasoning.
According to Artificial Analysis, four of the nine tests were first achieved by Carbon, including HLE-Diamund, Terminal-Bench-Science, AA-Omniscence and Automation Bench-AAA.
The tests covered difficult knowledge and reasoning, end-use operations in the scientific environment, factual knowledge and cross-tool automation tasks.
However, the fact that first on the Consolidated List does not mean that Carbon outperformed competitors on all the subdivisions.
In Artificial Analysis Commission Index, the Gemini 4 Carbon High Logic Configuration with Antigravity CLI received 59 points, while Claude Code with Claude Opus 5.5 High Logic Configuration obtained 62 points. Carbon ranked second and the gap narrowed to three.
Of these, in the FrontierCode v1.1 test, Carbon received 51% and Claude Opus 5.5 received 56%. This means that Anthropic still has an advantage in some of the real software engineering tasks.
On the other hand, in the AA-Briefcase benchmark for measuring expertise, Carbon received 1609 Elo, up 126 points from 1483 Elo in Argon. Among them, presentation missions were upgraded 293 Elo, analytical quality was enhanced 120 Elo.
However, Carbon still lags behind some of the competition models in this test, including Grok 4.7, Muse Wonder and Kimi K3.1 Pro.
In other words, Google has achieved the first comprehensive intelligence index, but in the sub-areas of programming AI agent, professional work and so forth, the lead position remains fiercely contested.
It is worth noting that Google also places special emphasis on Carbon ' s web development capacity in official announcements.
On WebDev Arena, the Elo of Carbon rose by 89 points to about 1768, ranked fourth. Google attributed this change to the improved quality of the results of the model-generated web pages, as well as to the development of a whole house, multimodular understanding and the enhancement of complex engineering capabilities.
From code generation to terminal execution, to complete software engineering tasks, the improvement of Carbon is fairly clear: It no longer seeks to generate a single correct code, but rather to enable the model to perform its tasks in a real development environment.
This change is also becoming the core of the current flagship model competition.
Google says carbon can work continuously. days
Over the past two years, the programming capacity of LLM has undergone significant changes.
Initially, the developers were concerned about whether the model could write a function, repair a bug, and then gradually move to allow AI to understand the code warehouse, call tools, perform tests and even deliver complete software items independently.
Today, the front-line model is facing a more complex problem: If a project mission takes hours or even days to complete, can AI continue with repeated failures, debugging and modification?
That's where Carbon's focus has been sharpened.
Google, in its official technical presentation, disclosed that in the internal engineering environment, researchers had observed that Carbon was running for several days. It can track hidden errors in large code libraries, analyse bottom system problems and attempt to develop operational restoration options.
Such missions usually do not have a clear solution. Developers may need to read a large number of codes, repeat problems, try different ways of repair, and adjust their judgement based on the results of the tests.
For AI agent, the real difficulty lies not only in generating codes but also in maintaining a long-term understanding of mission objectives, performance status and engineering constraints.
Google indicated that Carbon had optimized its work specifically for such long-range missions, emphasizing that models continued to analyse and find solutions after encountering a failed path, rather than simply ending the mission.
This capability orientation largely reflects the evolution of the AI programming tool.
In the traditional code complete scenario, developers can check the model output instantaneously, while in the long-range AI agent mission, models often require dozens or even hundreds of tool calls to be performed independently. Once an earlier judgement is miscalculated, subsequent implementation may gradually deviate from the original objective.
Thus, the ability of models to detect errors in a timely manner, reprogramme after failure, and identify their unfinished tasks becomes as important as programming capacity. From this point of view, the technical focus of Carbon is to try to get AI to complete a longer and more reliable working loop in complex engineering.
If complex programming is the ability of carbon's biggest developer, cybersecurity is probably the most strategic direction of Google's publish this time.
Google published some internal engineering data.
In the security code generation task, the success rate of Carbon is more than double that of Argon. Internal tests by the Crome team also show that Carbon has improved by more than 10 times the performance of a large front-line model from another laboratory in generating high-quality patches.
It's going up, but it's not going up.
While model performance continues to improve, Google chooses to let carbon follow Argon's initial pricing.
According to official announcements, Carbon's initial API price was $2 per million input token, $10 per million output token and 95 per cent discount for cache input.
But here's an easy-to-neglect detail: it's a first-off preferential price, not a permanent price.
In a footnote to the official announcement, Google made it clear that after the initial offer, Carbon would return to Token 4 per million and Token 20 per million. Google has not yet announced a specific expiry date for the Carbon concession period.
In terms of price alone, Carbon did not raise the unit cost of Token compared to Argon, but performance increased significantly.
However, token unit prices do not fully reflect the actual costs for developers who run complex AI agent missions.
Token, as well as tool call and running time, will affect overall costs.
The assessment of Artificial Analysis provides a more specific perspective. In its comprehensive intelligence index test, the average single mission cost of carbon was about $7.76 at standard prices, while Claude Opus 5.5 ' s top reasoning configuration was about $10.04.
This means that under this particular test, the average single mission cost of carbon is about 23 per cent lower than Claude Opus 5.5.
Carbon, can you win back the capital trust for Google?
It is quite dramatic that on the same day that Google publish Gemini 4 Carbon was retaken to the top of the global comprehensive list of LLM, its parent company, Alphabet, was hit by a cold in the capital market.
On the same day, model publish, Alphabet published the third quarter of 2026. The data show that the company collected $12,758 million in the current season, an increase of 25 per cent over the same period, but lower than expected by approximately $12,797 million on Wall Street; the net profit was $26,177 million, a decline of about 25 per cent over the same period. After the release of the financial statements, Alphabet stock prices were once down by more than 10 per cent in post-card transactions.
It is noteworthy that Google's AI business is not without growth. In the third quarter, Google Cloud received $30.71 billion, an increase of 103 per cent over the same period and a new growth record. Google CEO Sundar Pichai also revealed that Gemini applications have surpassed a billion users.
What is really worrying investors is the continued expansion of AI investment costs and when these inputs can be translated into sufficiently substantial profits.
According to Reuters, the free cash flow in Alphabet was negative in the third quarter of $9.07 billion, a further deterioration from the previous quarter of $5.9 billion. At the same time, the company projects that capital expenditure will increase significantly in 2027. At the same time, the growth in revenue from Google search operations slowed from 17 per cent in the previous quarter to 12 per cent, and the growth prospects for advertising are also being addressed.
Of course, Alphabet's decline in profits is not entirely attributable to AI investment. The loss of some $8.7 billion in equity investment was also recognized during the current season, with a clear impact on net profits and per share of earnings.
But it still reveals the current contradictions facing the AI industry.
On the one hand, Gemini 4 Carbon ' s publish demonstrates that Google still has the capacity to take the lead in front-line model competition; on the other hand, better-trained models, larger computing infrastructure, and maintenance of high-speed technological trajectories also mean higher and higher investment.
For Google, winning first on the LLM list and winning the trust of the capital market is becoming two different races.
References:
https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
https://artificialanalysis.ai/articles/gemini-4-carbon-new-number-1
Auto-fetched and translated by AI Atlas. All rights belong to the original author.