今天起 Gemini 4 Carbon 正式上线,但不是所有人都用得上Gemini 4 Carbon Officially Rolls Out Today, But Not for Everyone


AI free token/AI免费积分

AI 大模型的发布节奏已经快到让人应接不暇。

就在刚刚,Google DeepMind 正式宣布,旗下最新旗舰模型 Gemini 4 Carbon 正式面向开发者、企业、科研人员及普通消费者给出服务。

Carbon 是当下 Gemini 家族综合能力最强的公开上线模型,尤其擅长复杂编程、网络安全、科学推理,以及需要长时间执行的多智能体任务。

Image

图|官方博客🔗 https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/

两周前,谷歌已经借助 Fairwind 计划向部分网络安全防御人员开放 Carbon,如今进一步扩大了使用范围。

与此前发布的 Gemini 4 Argon 相比,Carbon 主要面向复杂软件工程、跨领域研究等高难度任务,可以消耗更多 Token 开展深入推理、反复验证和错误修正,而 Argon 继续承担大多数日常编程和知识工作。

Image

在 CursorBench 4.0、Perplexity WANDR 和 Zapier AutomationBench 三项基准测试中,Carbon 均排名第一,单任务成本比表现次优的外部竞争模型低了一半以上。

售价维度,Carbon 的 API 优惠售价与 Argon 相同,每百万输入 Token 为 2 美元,每百万输出 Token 为 10 美元,12 月 1 日起将分别调整至 4 美元和 20 美元。

Image

图|🔗 https://aistudio.google.com/docs/pricing

一句话召唤 AI 团队,连夜造出完整游戏

Gemini 4 Carbon 的一项关键升级,是可以协调多个 AI Agent,共同完成大型工程项目。

谷歌为此在 AI 编程平台 Google Antigravity 中推出了 /swarm 能力预览版,开发者输入一条指令,就能组织多个 AI Agent 分工执行任务,并实时查看进度。

其中,Carbon 可以负责架构外观、任务协调和代码审查,其他速度更快、成本更低的模型负责具体开发,还有专门的验证 Agent 检查修改结果。

谷歌展示的一款开放世界 3D 游戏,就由这样一支 AI 团队完成。

玩家扮演一只骑着自行车的鹈鹕,在谷歌山景城园区自由探索。开发过程涵盖 Blender 三维建模、Three.js 和 WebGL 游戏开发、程序化音效制作,以及使用 Lyria 生成背景音乐。

Image

另一项演示更加夸张。

谷歌向 Carbon 给出了 7 份公开的城市规划 PDF 文件,其中包含 206 张工程图纸,要求 AI 重建谷歌 Platform 37 办公楼。

Image

经过 10 轮通宵运行的多智能体任务,Carbon 最终制作出包含约 90.6 万个三角形的交互式 Three.js 三维模型。

Carbon 还依据埃菲尔铁塔建设期间的历史照片和工程图纸,制作出可交互的 3D 历史展览。

Image

甚至依据 CAD 图纸和 X 光影像,还原了一架具备 61 键机械结构的古钢琴。

Image

在网页外观维度,Carbon 仅借助两轮提示词,就为 Google Merchandise Store 提出了五套视觉外观思路,并利用 Nano Banana 2.1 生成商品图片。

Image

第三方开发者也分享了实际感受。

OpenCode 软件工程师 Kit Langton 表示,团队此前一直无法让桌面应用适配现有 UI 框架,于是在下班前向 Carbon 提交了一份需要清单。第二天回来时,Carbon 已经从零开发出一套 Rust UI 库,并重新编写了 OpenCode 桌面应用。

HubX 则在具备约 1.4 万项测试的代码库中,让 Carbon 处理 16 个真实 Bug,最终报告了 94% 的难题应对率,并且没有引入新的回归难题。

两小时发现零日漏洞,网络安全能力进一步提高

网络安全是 Gemini 4 Carbon 的另一项核心能力,在 SEC-bench Pro、ExploitGym 和 CWE-bench 等漏洞发现、利用和修复基准测试中,Carbon 均取得领先成绩。

在谷歌内部自主安全平台 Project Fortify 中,Carbon 的漏洞扫描召回率达到 100%,漏洞分类精确率达到 95%。在开源代码仓库测试中,其召回率相比 Argon 又取得超过 25% 的相对提高。

Image

更引人注意的是,在实际测试中,Gemini 4 Carbon 协助 Wiz 在不到两小时内发现两家大型 SaaS 服务商的零日漏洞,将原本数周的研究工作缩短至几小时。

谷歌云团队利用它验证了两个开源软件漏洞的可利用性,而 SentinelOne 则借助 Carbon 额外发现了 4 个其他参测模型未能识别的有效漏洞。

伴随网络安全能力提高,谷歌也加强了模型的安全防护。

在模拟网络攻击测试中,Carbon 成功阻止了自适应红队 Agent 发起的 99% 攻击路径,单次攻击成功突破防护的比例从 Argon 的 3.1% 降至 0.8%。

数周完成密码学验证,AI 蛋白质研究取得新突破

科学研究维度,谷歌公布了三个内部科研案例,分别涉及密码学形式化验证、生物实验分析和蛋白质结构预测。

其中最值得关注的,是 Signal 加密协议的代码验证。

谷歌研究人员开发了一套名为 AlphaProof Code 的智能体形式化验证框架,借助 Carbon 等模型将安全目标转化为数学定理,再生成数百万行 Lean 证明代码。

Image

最终,团队在数周内完成了 Signal 加密协议相关代码关键安全性质的数学证明,而类似工作过去通常需要专业团队花费数年。

在生物实验领域,Carbon 可以从实验记录和原始数据中识别异常,例如深孔板自动混合不充分、仪器中残留微量样品污染等,协助研究人员改善实验重复性。

蛋白质结构预测维度,谷歌利用 Carbon 组成多智能体平台,在大规模序列和结构数据库中自动寻找远缘同源蛋白,为 AlphaFold 3 等结构预测模型给出更多参考信息。

在 CASP17 盲测数据集中的 7 个低同源性蛋白质家族上,Carbon 找到的同源序列数量平均增加了 40 倍。相关成果有望协助研究人员在不重新训练结构预测模型的情况下,提高部分难研究蛋白质的预测能力。

从写代码到做财务分析,AI 开始承担更完整的工作

除了工艺研发,Carbon 也在向金融、法律和企业运营等专业领域扩展。

谷歌与 Artificial Analysis 合作推出了 Multimodal Knowledge Agent(MKA)基准测试,专门评估 AI Agent 处理 PDF、电子表格、图表及工程图纸的能力。

Carbon 在该测试中排名第二,但平均单任务成本比排名第一的模型低了六倍以上。

Image

在金融情形中,Carbon 在 Hebbia 的端到端金融 Agent 实测中追平了参测表现最好的模型,并在 Balyasny Asset Management 涵盖 23 种前沿模型规格的综合金融智能测试中名列前茅。

Hebbia 表示,Carbon 可以制作结构更清晰的 Excel 财务模型,以及质量更高的 PowerPoint 演示文稿。

企业搜索公司 Glean 则发现,Carbon 可以从办公数据中寻找指定的会议纪要、邮件或业务数据,并直接生成分析结果和行动建议。

非付费使用者什么时候能用上?

随着 Gemini 4 Carbon 正式开放,谷歌也公布了面向不同使用者的使用方式。

开发者可以借助 Google AI Studio 和 Android Studio 的 Gemini API 付费层调用 Carbon,企业使用者则可以借助 Gemini for business 及相关开发工具使用。

普通消费者维度,Gemini 4 Carbon 已开始面向 Google AI Ultra 订阅使用者开放,覆盖 Gemini App 和 Google 搜索 AI Mode。

Image

科研人员可以申请谷歌此前推出的免费研究计划,符合条件的参与者可取得为期六个月的 Google Antigravity 高使用额度访问权限,使用 Carbon 和 Argon 两款模型。网络安全研究人员则需要借助 Fairwind Program 申请专门的 Carbon for cyber。

回头看过去的消费电子市面,手机厂商的旗舰机型通常有着相对固定的发布周期,每年几场关键发布会,就足以让科技爱好者讨论上好几个月。

如今的 AI 行业却是另一番景象,几乎每天都有不同特色的新模型登场,有的刷新编程榜单,有的攻克数学难题,还有的已经可以连续工作几十个小时。

很难想象,笔者已经进入了一个模型多到感受不过来的时代。

如此密集的工艺更新,对使用者而言当然是件好事,只是看完 Gemini 4 Carbon 的发布公告,多少还有些意犹未尽。毕竟,谷歌展示了这么多令人眼馋的能力,普通消费者却需要订阅 Google AI Ultra 才能感受,多少有些不够意思。

本文来源:https://www.ifanr.com/1688883
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

The release pace of AI large models has become so fast that it is hard to keep up.

Just moments ago, Google DeepMind officially announced that its latest flagship model, Gemini 4 Carbon, is now officially available to developers, enterprises, researchers, and general consumers.

Carbon is the strongest publicly released model in the Gemini family to date in overall capability, especially strong at complex programming, cybersecurity, scientific reasoning, and long-running multi-agent tasks.

Image

Image | Official blog 🔗 https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/

Two weeks ago, Google had already opened Carbon to some cybersecurity defenders through the Fairwind program; now it has further expanded access.

Compared with the previously released Gemini 4 Argon, Carbon is mainly aimed at high-difficulty tasks such as complex software engineering and cross-domain research. It can consume more tokens to carry out deep reasoning, repeated verification, and error correction, while Argon continues to handle most everyday programming and knowledge work.

Image

In the three benchmarks CursorBench 4.0, Perplexity WANDR, and Zapier AutomationBench, Carbon ranks first, with the per-task cost more than half lower than that of the runner-up external competing models.

On pricing, Carbon's promotional API price is the same as Argon: $2 per million input tokens and $10 per million output tokens; starting December 1, these will be adjusted to $4 and $20 respectively.

Image

Image | 🔗 https://aistudio.google.com/docs/pricing

Summon an AI team in one sentence, and build a complete game overnight.

One key upgrade in Gemini 4 Carbon is that it can coordinate multiple AI agents to jointly complete large engineering projects.

For this, Google has introduced the /swarm capability preview in the AI coding platform Google Antigravity. With a single instruction, developers can organize multiple AI agents to divide tasks and check progress in real time.

Carbon can be responsible for architecture and design, task coordination, and code review; other faster, lower-cost models handle the concrete development, while dedicated verification agents check the results of the changes.

An open-world 3D game shown by Google was completed by such an AI team.

The player plays as a pelican riding a bicycle, freely exploring Google's Mountain View campus. The development process included Blender 3D modeling, Three.js and WebGL game development, procedural sound production, and the use of Lyria to generate background music.

Image

Another demo is even more over-the-top.

Google gave Carbon 7 publicly available urban planning PDF documents, containing 206 engineering drawings, and asked the AI to reconstruct Google’s Platform 37 office building.

Image

After 10 rounds of overnight multi-agent tasks, Carbon finally produced an interactive Three.js 3D model containing about 906,000 triangles.

Carbon also created an interactive 3D historical exhibition based on historical photos and engineering drawings from the construction of the Eiffel Tower.

Image

It even restored a historical piano with a 61-key mechanical structure based on CAD drawings and X-ray images.

Image

In terms of webpage visual design, Carbon proposed five visual direction concepts for the Google Merchandise Store using only two rounds of prompts, and used Nano Banana 2.1 to generate product images.

Image

Third-party developers also shared their real-world experience.

OpenCode software engineer Kit Langton said the team had long been unable to make their desktop app work with existing UI frameworks, so before leaving work they submitted a requirements list to Carbon. When they came back the next day, Carbon had developed a Rust UI library from scratch and rewrote the OpenCode desktop app.

HubX had Carbon handle 16 real bugs in a codebase with about 14,000 tests, and ultimately reported a 94% success rate on tough issues, without introducing new regression issues.

Discovering zero-day vulnerabilities in two hours, cybersecurity capabilities further improved

Cybersecurity is another core capability of Gemini 4 Carbon, which achieved leading results in benchmarks for vulnerability discovery, exploitation, and remediation such as SEC-bench Pro, ExploitGym, and CWE-bench.

In Google’s internal autonomous security platform Project Fortify, Carbon achieved a 100% vulnerability scanning recall and 95% vulnerability classification precision. In open-source repository testing, its recall achieved a relative improvement of over 25% compared with Argon.

Image

Even more notable is that, in real-world testing, Gemini 4 Carbon helped Wiz discover zero-day vulnerabilities at two large SaaS providers in under two hours, cutting research work that would typically take weeks down to a few hours.

The Google Cloud team used it to verify the exploitability of two open-source software vulnerabilities, while SentinelOne leveraged Carbon to additionally discover four valid vulnerabilities that other evaluated models failed to identify.

As cybersecurity capabilities improved, Google also strengthened the model’s security protections.

In simulated cyberattack testing, Carbon successfully blocked 99% of attack paths launched by adaptive red-team agents, reducing the proportion of single attacks that successfully breached defenses from 3.1% for Argon to 0.8%.

Cryptographic verification completed in weeks, new breakthrough achieved in AI protein research

In the scientific research dimension, Google announced three internal research cases involving cryptographic formal verification, biological experiment analysis, and protein structure prediction.

Among them, the most noteworthy is the code verification of the Signal encryption protocol.

Google researchers developed an agent formal verification framework named AlphaProof Code, leveraging models like Carbon to transform security goals into mathematical theorems, and then generating millions of lines of Lean proof code.

Image

Ultimately, the team completed the mathematical proof of key security properties for the Signal encryption protocol's related code within weeks, whereas similar work typically required professional teams years in the past.

In the field of biological experiments, Carbon can identify anomalies from experimental records and raw data, such as insufficient automated mixing in deep-well plates or trace sample contamination remaining in instruments, assisting researchers in improving experimental reproducibility.

In the dimension of protein structure prediction, Google utilized Carbon to form a multi-agent platform that automatically searches for distant homologous proteins in large-scale sequence and structure databases, providing more reference information for structure prediction models like AlphaFold 3.

On 7 protein families with low homology in the CASP17 blind test dataset, the number of homologous sequences found by Carbon increased by an average of 40 times. The related results are expected to assist researchers in improving the prediction capabilities of some difficult-to-study proteins without retraining structure prediction models.

From writing code to financial analysis, AI begins to take on more complete tasks

Beyond process research and development, Carbon is also expanding into professional fields such as finance, law, and enterprise operations.

Google collaborated with Artificial Analysis to launch the Multimodal Knowledge Agent (MKA) benchmark, specifically evaluating AI Agents' ability to process PDFs, spreadsheets, charts, and engineering drawings.

Carbon ranked second in this test, but its average cost per task was more than six times lower than the top-ranked model.

Image

In financial scenarios, Carbon matched the best-performing model in Hebbia's end-to-end financial Agent tests, and ranked among the top in Balyasny Asset Management's comprehensive financial intelligence tests covering 23 cutting-edge model specifications.

Hebbia stated that Carbon can create Excel financial models with clearer structures, as well as higher-quality PowerPoint presentations.

Enterprise search company Glean found that Carbon can locate specified meeting minutes, emails, or business data from office data, and directly generate analysis results and actionable recommendations.

When will non-paying users be able to use it?

As Gemini 4 Carbon officially opens, Google has also announced usage methods for different users.

Developers can access Carbon via the paid tier of the Gemini API in Google AI Studio and Android Studio, while enterprise users can access it through Gemini for business and related development tools.

From a general consumer perspective, Gemini 4 Carbon has started rolling out to Google AI Ultra subscribers, available via the Gemini App and Google Search AI Mode.

Image

Researchers can apply for the free research program previously launched by Google; eligible participants will receive six months of access to high usage quotas for Google Antigravity, allowing them to use both the Carbon and Argon models. Cybersecurity researchers, however, need to apply for the specialized Carbon for cyber through the Fairwind Program.

Looking back at the consumer electronics market, flagship models from phone manufacturers typically followed a relatively fixed release cycle; a few key launch events each year were enough to keep tech enthusiasts discussing for months.

Today, the AI industry presents a different scene: new models with distinct features debut almost every day. Some top programming leaderboards, others solve complex math problems, and some can already work continuously for dozens of hours.

It is hard to imagine that I have entered an era where there are too many models to keep up with.

Such intensive updates are certainly a good thing for users, but after reading the announcement of Gemini 4 Carbon, I still feel somewhat unsatisfied. After all, Google has showcased so many coveted capabilities, yet ordinary consumers need to subscribe to Google AI Ultra to experience them, which feels a bit stingy.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry