AI 模型和平台
分阶段推出发送 Google 的 Gemini 4 碳给网络捍卫者第一
在 Google 上的首选来源中添加 Unite.AI
Google 宣布 Gemini 4 Carbon在2026年10月28日将其描述为公司最强的前沿模型，称将通过分阶段的推出，从一个网络维权者子集开始，而不是广泛的发布，到达用户。
在公告中，Koray Kavukcuoglu,Google DeepMind的SVP和Google首席AI建筑师将Carbon定性为公司最强的模型，拥有最先进网络安全和长视距能力。 碳继承了Gemini 4 Argon，这是在碳公告前两周的Google 发布。 Google 称新模式显示了它推进前沿的步伐，同时加强保障。
Google 称正采取审慎的部署方式，并积极参与美国政府的自愿程序，进行释放前模型测试。 从宣布的一周开始，该公司将向一个网络维权者子集提供Carbon，然后通过它的Fairwind Program逐步逐步接触更多的信任的伙伴。 Google 称正在继续收集反馈，并在护栏上进行移动，然后通过产品表面向开发者、企业和消费者提供模型，并补充说它打算尽快这样做。
长线编码和基准测试将军
Google 称 Carbon是针对更困难，长视距的工程问题设计，为持续的代理工作流程构建的，内部团队已经看到该模型运行了数天，并在复杂，多步骤的任务中坚持通过死端，如在大型代码库和低层系统上修补细微的bug.
Google表示，Carbon在最近更新的人工分析智能指数5.0中排名第一，得分59. 公司报告称，Carbon在WebDev Arena上以1768分的得分获得了89elo的分数，达到第4名，他列举了网页开发产出和文物质量的提高。 Google 称的改进重点是编码、代理和科学任务，以及基于用户反馈的全网开发和多式联运。
网络安全防御性能
Google报告最先进在脆弱性检测、验证和补救的网络安全防御生命周期中取得了结果，称的外部评价没有护卫。 在SEC-bench Pro(测量现实世界软件复杂缺陷的自主发现)上，Google报告为88.2%. 在勘探Gym上，一个基准测试用来验证已查明的弱点是否可以开采，Google报告新的最先进得分为54.8%. Google报告为71.0%。
在内部的基准测试上，Google报告说，Carbon在生成安全码在Argon上空的成功率上实现了2x以上的提升，其Chrome团队发现Carbon在生成高质量补丁方面比另一个实验室的大型前沿模型做得更好。 Google补充说，Wiz，它的云安全平台，观测到在基准测试上领先的Carbon进行网络渗透测试，这是Wiz扫描Good倡议的核心能力。
安全与协调成果
碳基于用于Argon的多层次安全保障堆，该堆结合了模型训练，实时滥用防御，即时注射韧性，以及思维链监测，根据Google.
公司报告称，Carbon将终端-Bench 4.0成功作弊率降至0.0 %，作弊尝试降至0.4%，同时整体准确度提高，在与内部团队的直播编码环境中，高高度错配预警下降了近80%. 在内部测试中，Google 称，Carbon 削减在Argon上空的大约2x不支持声称，直接将方法上的告诫放入可交付品中，并扣下有明确清单的仍然需要的行动，而不是围绕缺失的证据编造或套期。 Google还报告说，在多步骤计划执行期间，超覆盖率减少26%，服务不足率减少29%，称模型在任务模糊时采取行动前，会向用户检索相关来源并显示差异。 在Gray Swan的间接提示词注入基准测试上，Google报告说，在Argon之前以0.7%领导基准测试之后，Carbon——抵制每一次评估的攻击尝试——的攻击成功率为0.0 %.
Fairwind 程序访问和定价
由Google DeepMind运营的Fairwind方案，给予政府、保健提供者和电信服务等高度优先维护者早期获得先进模型的机会。 Google DeepMind表示，它目前与全球超过650个合作伙伴合作，一组程序合作伙伴获得Gemini 4 Argon的独家访问权，与CodeMender使用，其专用代码安全agent用于软件自动修正。
方案准入有具体的条件。 合作伙伴只能执行允许的双重用途任务，包括授权的威胁模拟、逆向工程和恶意软件分析，用于防御和学术研究目的；不允许制造恶意软件等恶意任务。 参与组织必须采用用户级认证和耐网信多要素认证，限制模式进入内部网络安全，事件应对，或渗透测试团队，跟踪员工访问和使用。 合作伙伴不得分享、重新分配或出售对模型的获取，Google DeepMind对申请人进行背景调查，以核实安全历史和道德操作记录。 该方案优先安排政府和国家网络当局、关键基础设施运营商跨越保健、电信、能源和金融网络以及核心技术平台。
碳发布的引入价格与Argon相同：每百万输入令牌2元，每百万输出令牌10元，其中缓存输入令牌的价格为输入令牌价格的95%. 公告的一个脚注指出，在导言期结束后，定价为每百万美元输入令牌4美元，每百万美元产出令牌20美元。
Google 称急切希望看到最初的网络维权者团体如何使用这个模式
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。
AI Models & Platforms
Staged Rollout Sends Google’s Gemini 4 Carbon To Cyber Defenders First
Add Unite.AI to your preferred sources on Google
Google announced Gemini 4 Carbon on October 28, 2026, describing it as the company’s strongest frontier model, and said it will reach users through a staged rollout that begins with a subset of cyber defenders rather than a broad release.
In the announcement, Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google, characterized Carbon as the company’s strongest model, with state-of-the-art cybersecurity and long-horizon capabilities. Carbon succeeds Gemini 4 Argon, which Google launched two weeks before the Carbon announcement. Google said the new model illustrates the pace at which it is pushing the frontier forward while strengthening its safeguards in tandem.
Google said it is taking a deliberate approach to deployment and is actively engaged in the U.S. government’s voluntary process for pre-release model testing. Starting the week of the announcement, the company will make Carbon available to a subset of cyber defenders, then gradually phase access to additional trusted partners through its Fairwind Program. Google said it is continuing to gather feedback and iterate on guardrails before making the model available to developers, businesses, and consumers through its product surfaces, adding that it intends to do so as soon as possible.
Long-Horizon Coding and General Benchmarks
Google said Carbon is designed for harder, long-horizon engineering problems and built for sustained agentic workflows, and that internal teams have seen the model run for days and persist through dead ends on complex, multi-step tasks such as root-causing and patching subtle bugs across large codebases and low-level systems.
According to Google, Carbon ranks first on the recently updated version 5.0 of the Artificial Analysis Intelligence Index with a score of 59. The company reported that Carbon gained 89 elo points to reach fourth place with a score of 1768 on WebDev Arena, citing improved quality of web development output and artifacts. Google said it focused its improvements on coding, agentic, and scientific tasks, as well as full-stack web development and multimodal understanding, building on user feedback.
Cybersecurity Defense Performance
Google reported state-of-the-art results across the cybersecurity defense lifecycle of vulnerability detection, validation, and remediation, and said the external evaluations were run without guardrails. On SEC-bench Pro, which measures the autonomous discovery of complex flaws in real-world software, Google reported 88.2%. On ExploitGym, a benchmark used to validate whether identified vulnerabilities are exploitable, Google reported a new state-of-the-art score of 54.8%. On Collinear’s CWE-bench, which tests patching vulnerabilities at the root cause, Google reported 71.0%.
On internal benchmarks, Google reported that Carbon achieves more than 2x uplift in success rate for generating secure code over Argon, and that its Chrome team found Carbon performed more than 10X better than another lab’s large frontier model at generating high-quality patches. Google added that Wiz, its cloud security platform, observes Carbon leading on benchmarks for web penetration testing, a core capability for Wiz’s Scan for Good initiative.
Safety and Alignment Results
Carbon builds on the multi-layered safety and security stack used for Argon, which combines model training, real-time misuse defenses, prompt-injection resilience, and chain-of-thought monitoring, according to Google.
The company reported that Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, with cheating attempts falling to 0.4% while overall accuracy improved, and that high-severity misalignment alerts dropped by almost 80% in live coding environments with internal teams. In internal testing, Google said, Carbon cuts unsupported claims by approximately 2x over Argon, places methodological caveats directly into deliverables, and withholds blocked actions with a clear list of what remains required rather than fabricating or hedging around missing evidence. Google also reported reductions of 26% in overreach and 29% in under-service during multi-step plan execution, and said the model retrieves relevant sources and surfaces discrepancies to the user before acting when tasks are ambiguous. On Gray Swan’s Indirect Prompt Injection Benchmark, Google reported a 0.0% attack success rate for Carbon — resisting every evaluated attack attempt — after Argon previously led the benchmark at 0.7%.
Fairwind Program Access and Pricing
The Fairwind Program, run by Google DeepMind, gives high-priority defenders such as governments, healthcare providers, and telecommunications services early access to advanced models. Google DeepMind states that it currently works with more than 650 partners globally, and that a set of program partners receives exclusive access to Gemini 4 Argon for use with CodeMender, its specialized code security agent for automating software fixes.
Program access carries specific conditions. Partners may carry out only permitted dual-use tasks, including authorized threat simulation, reverse engineering, and malware analysis for defensive and academic research purposes; malicious tasks such as creating malware are not permitted. Participating organizations must apply user-level authentication and phishing-resistant multi-factor authentication, limit model access to internal cybersecurity, incident response, or penetration testing teams, and track employee access and use. Partners may not share, redistribute, or sell access to the models, and Google DeepMind conducts background checks on applicants to verify security history and a record of ethical operations. The program prioritizes governments and national cyber authorities, critical infrastructure operators across healthcare, telecommunications, energy, and financial networks, and core technology platforms.
Carbon launches at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off the input token price. A footnote in the announcement states that after the introductory period expires, pricing will be $4 per million input tokens and $20 per million output tokens.
Google said it is eager to see how the initial cohort of cyber defenders uses the model and looks forward to bringing Carbon to developers, businesses, and consumers soon.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.