原文刊于 Google The Keyword 博客，作者 Koray Kavukcuoglu（Google DeepMind 高级副总裁、Google 首席 AI 架构师）
Gemini 4 Carbon is now available
Gemini 4 Carbon 现已推出
Developers, enterprises, scientists, and consumers can now try our latest frontier model.
开发者、企业、科学家和普通用户，现在都可以试用我们最新的前沿模型。
Nov 10, 2026
Koray Kavukcuoglu
SVP, Google DeepMind and Chief AI Architect, Google
2026年11月10日（官方日期；发布时刻为 16:00 UTC，即 亚洲/上海 11月11日 00:00）
Koray Kavukcuoglu
Google DeepMind 高级副总裁、Google 首席 AI 架构师
Today, we’re making Gemini 4 Carbon available to developers, enterprises, scientists, and consumers. Carbon is our strongest model with frontier performance in cybersecurity capabilities and long-horizon agentic workflows across complex knowledge work, software engineering, and scientific reasoning.
今天，我们向开发者、企业、科学家和普通用户开放 Gemini 4 Carbon。Carbon 是我们最强的模型，在网络安全能力，以及跨越复杂知识工作、软件工程和科学推理的长时程智能体工作流上，都达到了前沿水平。
Two weeks ago, we released Carbon to a set of cybersecurity defenders via our Fairwind Program. Following our recent launch of Gemini 4 Argon, Carbon pushes the frontier even further. Argon remains the ideal choice for the majority of software engineering and knowledge work, while Carbon is built for tasks that test the limits of a model’s intelligence like complex, multi-step coding and debugging and open-ended scientific research and discovery — using more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops.
两周前，我们通过 Fairwind 计划，向一批网络安全防御者发布了 Carbon。在最近推出 Gemini 4 Argon 之后，Carbon 把前沿又向前推进了一步。对大多数软件工程和知识工作而言，Argon 仍然是更合适的选择；而 Carbon 则专为考验模型智力极限的任务而构建，例如复杂的多步骤编码与调试，以及开放式的科学研究与发现——它会比 Argon 使用更多 token，以便更充分地探索、更深入地推理，并运行更长的验证循环。
Across independent benchmarks designed to test model performance in real workflows, Carbon pairs frontier accuracy with practical production economics. It ranks #1 on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench at more than 2X lower cost per task compared to the next best external model.
在一批用于检验模型在真实工作流中表现的独立基准上，Carbon 把前沿精度与实际可用的生产成本结合在一起。它在 CursorBench 4.0、Perplexity 的 WANDR 和 Zapier 的 AutomationBench 上排名第一，且每项任务的成本不到次优外部模型的一半。
To put Carbon’s capabilities to work on larger and complex projects, we’re previewing /swarm in Google Antigravity, turning a single prompt into a coordinated team of agents you can track in real time. Developers can assign Carbon to roles that demand the deepest reasoning, like architectural design, coordination, and auditing, while pairing it with faster, lighter models across parallel isolated workspaces and adversarial verifiers that test every change.
为了把 Carbon 的能力用到更大、更复杂的项目上，我们正在 Google Antigravity 中预览 /swarm：一条提示词就能变成一支可实时跟踪的协同 Agent 团队。开发者可以把 Carbon 分配给最需要深度推理的角色，例如架构设计、协调和审计，同时在并行的隔离工作区中搭配更快、更轻量的模型，以及用于检验每一处改动的对抗式验证器。
Gemini 4 Carbon is available via the API in Google AI Studio and Gemini for business for all paying developers and enterprises, at the same introductory price1 as Gemini 4 Argon ($2 per million input tokens and $10 per million output tokens). It is also rolling out to all Google AI Ultra subscribers in Google Antigravity, the Gemini app, and AI Mode in Search.
对于所有付费开发者和企业，Gemini 4 Carbon 已可通过 Google AI Studio 和 Gemini for business 中的 API 使用，定价与 Gemini 4 Argon 的导入期价格相同1
（每百万输入 token 2 美元，每百万输出 token 10 美元）。它也正在向所有 Google AI Ultra 订阅用户推出，可在 Google Antigravity、Gemini 应用 以及 Search 中的 AI Mode 使用。
Higher-precision knowledge work, artifact generation, and data analysis更高精度的知识工作、产物生成与数据分析
Across legal, finance, and enterprise operations, Gemini 4 Carbon combines frontier visual understanding, reasoning, and exhaustive retrieval to synthesize grounded insights across large sets of dense documents and data. It follows strict schemas and formatting to generate polished and precise spreadsheets, presentations, and reports.
在法律、金融和企业运营场景中，Gemini 4 Carbon 结合前沿的视觉理解、推理和穷尽式检索，从大量密集的文档和数据中综合出有依据的洞察。它会遵循严格的模式和格式要求，生成精致、精确的电子表格、演示文稿和报告。
Existing benchmarks don’t adequately capture how agents reason natively over dense files with charts, images, blueprints, schematics, and tables, so we partnered with Artificial Analysis to launch the new Multimodal Knowledge Agent (MKA) benchmark that tests a model’s multimodal performance in agentic workflows. Carbon ranks #2 while being more than 6X lower in average cost per task compared to the top performing model.
现有基准还不足以衡量 Agent 如何原生地处理带有图表、图像、蓝图、示意图和表格的密集文件，因此我们与 Artificial Analysis 合作，推出了新的 Multimodal Knowledge Agent（MKA）基准，用于测试模型在智能体工作流中的多模态表现。Carbon 排名第 2，同时平均每项任务成本不到表现最好模型的六分之一。
Using Google Antigravity’s new /swarm feature, a team of Gemini 4 Carbon agents turned historical photos and engineering plates into an interactive Three.js 3D exhibition of the Eiffel Tower's 1887–1889 construction from a single prompt. Watch the video
借助 Google Antigravity 的新功能 /swarm，一组 Gemini 4 Carbon Agent 仅凭一条提示词，就把历史照片和工程图板，变成了一场关于埃菲尔铁塔 1887–1889 年建造过程的交互式 Three.js 3D 展览。观看视频
With seven public City Council planning PDFs (206 extracted engineering drawing views) and a single prompt, a multi-agent Carbon team using Google Antigravity’s /swarm feature built an interactive 906K-triangle Three.js reconstruction of Google’s Platform 37 office across 10 overnight loops. Watch the video
仅凭 7 份公开的市议会规划 PDF（从中提取出 206 张工程图纸视图）和一条提示词，一支使用 Google Antigravity /swarm 功能的多 Agent Carbon 团队，经过 10 个通宵循环，构建了 Google Platform 37 办公楼的交互式 Three.js 重建，共包含 90.6 万个三角形。观看视频
In Google Antigravity, Gemini 4 Carbon built an interactive multiscale textile visualizer based on an academic paper, modeling a virtual-fiber finite element framework with interactive 3D yarn meshes, 110-fiber polar tracking, and dynamic stress-strain curves with minimal prompting. Watch the video
在 Google Antigravity 中，Gemini 4 Carbon 基于一篇学术论文，构建了一个交互式多尺度纺织可视化工具，以极少的提示建模了虚拟纤维有限元框架，包括交互式 3D 纱线网格、110 根纤维的极坐标跟踪，以及动态应力-应变曲线。观看视频
In pre-release evaluations, Carbon matched the performance of the best model tested on deep financial research on Hebbia’s E2E Financial Agentic evaluation, and ranked at the top out of 23 frontier model configurations on Balyasny Asset Management’s General Financial Intelligence benchmark, spanning legal expertise, agentic tool use and execution, coding and data engineering, and perception and context understanding.
在发布前评估中，Carbon 在 Hebbia 的 E2E Financial Agentic 评测上，于深度金融研究任务中追平了所测最好模型的表现；并在 Balyasny Asset Management 的 General Financial Intelligence 基准上，在 23 种前沿模型配置中名列前茅。该基准涵盖法律专业能力、智能体工具使用与执行、编码与数据工程，以及感知与上下文理解。
Gemini 4 Carbon is the strongest model we’ve evaluated for deep financial research, building cleaner, better-structured Excel models and producing sharper PowerPoint decks. It is a real step forward for diligence workflows.
— Shanks Salwan, Lead Applied AI Engineer, Hebbia
Gemini 4 Carbon 是我们评估过的、最擅长深度金融研究的模型，能构建更干净、结构更好的 Excel 模型，并做出更锐利的 PowerPoint 演示文稿。这对尽职调查工作流而言，是一次真正的进步。
— Shanks Salwan，Lead Applied AI Engineer, Hebbia
Our benchmark is designed to look beyond general capability and evaluate the work financial teams actually do. Carbon stood out on tasks involving complex context and structured deliverables, especially charts, code, and tables.
— Audrey Liu, Senior AI Engineer, Balyasny Asset Management
我们的基准旨在超越一般能力评估，去衡量金融团队实际在做的工作。Carbon 在涉及复杂上下文和结构化交付物的任务上表现突出，尤其是图表、代码和表格。
— Audrey Liu，Senior AI Engineer, Balyasny Asset Management
Gemini 4 Carbon goes beyond research. It finds the call summary, the buried email or the exact pipeline figure across workplace data, then turns it into targeted analysis and next steps. In our evaluations, it led every frontier model we tested on task completion, and we're excited to bring it to Glean customers tackling complex, multi-step work.
— Thai Tran, AI Product Lead, Glean
Gemini 4 Carbon 的能力不只停留在研究。它能在工作场所数据中找到通话摘要、被埋没的邮件，或精确的管线数字，再把它们转化成有针对性的分析和下一步行动。在我们的评测中，它在任务完成方面领先于我们测试过的所有前沿模型。我们很期待把它带给那些正在处理复杂、多步骤工作的 Glean 客户。
— Thai Tran，AI Product Lead, Glean
Improved agentic coding, full stack development, and 3D and UI design更强的智能体式编程、全栈开发以及 3D 与 UI 设计
Carbon performs exceptionally well on engineering tasks that require multi-agent collaboration like multi-package migrations, cross-subsystem root-cause investigations, and complex architectural redesigns. For web app development, it runs build-and-verify loops, recovers from errors, and checks both its code and the rendered UI to produce complete, working applications. Stronger visual design, spatial reasoning, and attention to environmental detail also carry over to 3D game development and creative coding.
在需要多 Agent 协作的工程任务上，Carbon 表现尤为出色，例如多软件包迁移、跨子系统的根因调查，以及复杂的架构重新设计。在 Web 应用开发中，它会运行构建并验证的循环，从错误中恢复，并同时检查代码和渲染后的 UI，以产出完整、可运行的应用。更强的视觉设计、空间推理，以及对环境细节的关注，也延伸到了 3D 游戏开发和创意编程中。
When we struggled to get our app looking consistent with an existing UI framework, we gave Carbon a list of requirements before closing our day. Overnight, Carbon had built a new Rust UI library from scratch and rewritten our OpenCode desktop app with working demos. It’s an incredibly brainy model and now a preferred planning model for our hardest tasks.
— Kit Langton, Software Engineer, OpenCode
当我们难以让应用与现有 UI 框架保持外观一致时，我们在下班前给 Carbon 列了一份需求。一夜之间，Carbon 从零构建了一个新的 Rust UI 库，并重写了我们的 OpenCode 桌面应用，还带有可运行的演示。它是一个极其聪慧的模型，现在已成为我们处理最困难任务时优先选用的规划模型。
— Kit Langton，Software Engineer, OpenCode
Gemini 4 Carbon is by far the best frontier model for difficult cloth simulation. Its understanding of 3D, physics, and difficult (subtle) scientific decisions really make this model reliable. We spent 6 days working with it with minimal steering and it improved an existing, well tuned, soft body simulation in both speed and fidelity.
— Andrew Carr, Chief Scientist and Co-Founder, Cartwheel
在困难的布料仿真方面，Gemini 4 Carbon 显然是目前最好的前沿模型。它对 3D、物理，以及那些困难（且微妙）的科学决策的理解，让这个模型非常可靠。我们只做了极少的引导，就和它一起工作了 6 天，它在速度和保真度上都改进了一个已有的、调校良好的软体仿真。
— Andrew Carr，Chief Scientist and Co-Founder, Cartwheel
Gemini 4 Carbon treats legacy software contracts with absolute respect. Across 16 real bug fixes in our 14,000-test codebase, Carbon achieved a 94% resolution rate with zero regressions.
— Baki Celtik, AI Engineer, HubX
Gemini 4 Carbon 对遗留软件契约保持了绝对尊重。在我们拥有 14,000 项测试的代码库中，针对 16 个真实缺陷修复，Carbon 的解决率达到 94%，且零回归。
— Baki Celtik，AI Engineer, HubX
Our game characters must listen, follow world rules, and execute multi-step plans in real time. In testing, Carbon excelled at this. It accurately followed complex player requests and held firm world boundaries, explaining why impossible actions were disallowed. That kind of judgment makes characters feel truly alive.
— Brian Tanner, Co-Founder & CEO, Artificial Agency
我们的游戏角色必须倾听、遵守世界规则，并实时执行多步骤计划。在测试中，Carbon 在这一点上表现出色。它能准确遵循复杂的玩家请求，并坚守明确的世界边界，解释为什么不允许不可能的动作。这种判断力，让角色真正显得栩栩如生。
— Brian Tanner，Co-Founder & CEO, Artificial Agency
In our testing, Gemini 4 Carbon grounded its analysis in source content, adjusted its conclusions as evidence changed, and delivered clear recommendations in completed reviews. These capabilities are particularly valuable in highly regulated environments where decisions depend on the careful review of complex information. For example, Carbon scored 100% on public sector tasks, reinforcing our confidence in its ability to deliver accurate recommendations.
— Yashodha Bhavnani, VP of AI Products, Box
在我们的测试中，Gemini 4 Carbon 会把分析建立在源内容之上，随着证据变化调整结论，并在完成的审查中给出清晰建议。这些能力在高度受监管的环境中尤其有价值，因为那里的决策依赖于对复杂信息的审慎审查。例如，Carbon 在公共部门任务上得到了 100% 的分数，进一步增强了我们对其能够给出准确建议的信心。
— Yashodha Bhavnani，VP of AI Products, Box
Using Google Antigravity’s /swarm feature, Carbon built an open-world 3D game where you explore Google’s Mountain View campus as a pelican riding a GBike — coordinating subagents to model 3D assets in Blender from Omni and Nano Banana references, build the browser game in Three.js with WebGL shaders, and pair procedural Web Audio sound effects with Lyria music. Try the live app. Watch the video
借助 Google Antigravity 的 /swarm 功能，Carbon 构建了一款开放世界 3D 游戏：你可以化身一只骑着 GBike 的鹈鹕，探索 Google 山景城园区——它协调多个子 Agent，根据 Omni 和 Nano Banana 的参考在 Blender 中建模 3D 资产，用 Three.js 和 WebGL 着色器构建浏览器游戏，并把程序化的 Web Audio 音效与 Lyria 音乐搭配在一起。试试这个在线应用。观看视频
Paired with improved front end taste and motion design skills in Google Antigravity, Carbon reimagined the Google Merchandise Store in two prompts, proposing five art direction concepts and generating all-new product imagery with Nano Banana 2.1. Try the live app and Gemini front end design skill. Watch the video
配合 Google Antigravity 中更强的前端审美和动效设计能力，Carbon 只用两条提示词就重新构想了 Google Merchandise Store，提出了五种艺术方向概念，并用 Nano Banana 2.1 生成了全新的产品图像。试试这个在线应用和 Gemini 前端设计 skill。观看视频
Starting from a video reference generated with Gemini Omni in Google Antigravity, built a cinematic multi-shot underwater Atlantis scene. The pipeline modeled procedural ruins, volumetric caustic lighting, and water simulation in Blender before converting the full environment to Three.js for real-time browser rendering. Watch the video
从 Google Antigravity 中由 Gemini Omni 生成的视频参考出发，（原文此处未写出主语）构建了一个电影感的多镜头水下亚特兰蒂斯场景。这条流水线先在 Blender 中建模程序化遗迹、体积焦散光照和水模拟，再把完整环境转换到 Three.js，以便在浏览器中实时渲染。观看视频
Using a multi-agent loop in Google Antigravity, Carbon reconstructed a playable c. 1780 Stein-Schule fortepiano with 61-key escapement physics from CAD drawings and X-ray radiographs and placed it in a 3D model of the Royal Conservatory of Music's Koerner Hall in Toronto with seat-specific spatial acoustics. Try the live app. Watch the video
通过 Google Antigravity 中的多 Agent 循环，Carbon 根据 CAD 图纸和 X 光射线照片，重建了一架可演奏的约 1780 年 Stein-Schule 古钢琴，具备 61 键擒纵机械物理；并把它放进多伦多皇家音乐学院 Koerner Hall 的 3D 模型中，带有按座位区分的空间声学。试试这个在线应用。观看视频
Frontier cybersecurity capabilities前沿网络安全能力
Gemini 4 Carbon for cyber delivers our strongest cybersecurity capabilities to date, giving security teams the end-to-end intelligence they need in protecting critical software. When accessed through the Fairwind Program, Carbon leads on key benchmarks across finding (SEC-bench Pro), exploiting (ExploitGym), and patching (CWE-bench v1) vulnerabilities: inspecting software, tracing how flaws can be chained together across systems, and generating verified root-cause patches.
面向网络防御的 Gemini 4 Carbon for cyber，提供了我们迄今最强的网络安全能力，让安全团队在保护关键软件时拥有所需的端到端智能。通过 Fairwind 计划 使用时，Carbon 在发现（SEC-bench Pro）、利用（ExploitGym）和修补（CWE-bench v1）漏洞的关键基准上领先：它能检查软件，追踪缺陷如何在系统之间被串联起来，并生成经过验证的根因补丁。
Early testers inside and outside of Google are already seeing strong results with Carbon for cyber in practice.
Google 内外的早期测试者，已经在实践中看到 Carbon for cyber 的出色结果。
• In Project Fortify, Google’s autonomous system spanning scanning, triage, and remediation, Carbon hit 100% recall on scanning and 95% precision on triage, on our internal benchmark. On Open Source code repository testing we see an over 25% relative improvement on Argon's already strong recall rate, surfacing previously undetected, real-world logic flaws. Bringing these gains across Google, one of the world’s largest users of open-source software, helps secure both our own systems and the broader ecosystem.
在 Project Fortify（Google 覆盖扫描、分诊和修复的自主系统）中，Carbon 在我们的内部基准上，扫描召回率达到 100%，分诊精确率达到 95%。在开源代码仓库测试中，我们看到它在 Argon 已经很强的召回率基础上，又有超过 25% 的相对提升，并发现了此前未被检出的真实逻辑缺陷。把这些提升推广到 Google——全球最大的开源软件使用者之一——有助于同时保护我们自己的系统和更广泛的生态系统。
• Google’s Cloud Vulnerability Research (CVR) team recently used Carbon to prove that two bugs in a popular open-source software were in fact exploitable vulnerabilities, comparable in severity to the recent PHP2Shell exploit. By using Carbon to build a proof of concept, CVR helped maintainers re-classify the bug as critical and swiftly issue a patch. Details will be shared after the responsible disclosure process completes.
Google 的 Cloud Vulnerability Research（CVR）团队最近使用 Carbon 证明，一款流行开源软件中的两个缺陷实际上是可被利用的漏洞，严重程度可与近期的 PHP2Shell 利用 相比。CVR 借助 Carbon 构建概念验证，帮助维护者将该缺陷重新定为严重级别，并迅速发布了补丁。细节将在负责任披露流程完成后再公布。
• As part of Wiz’s Scan for Good program, the team used Carbon to assess vulnerabilities in public digital infrastructure. For example, when scanning two major SaaS providers, the model uncovered 0-day vulnerabilities in under two hours compressing weeks of vulnerability research to a single afternoon. The team partnered with the impacted organizations to remediate in record time, helping to mitigate potential impact on tens of thousands of companies.
作为 Wiz 的 Scan for Good 计划的一部分，该团队使用 Carbon 评估公共数字基础设施中的漏洞。例如，在扫描两家大型 SaaS 提供商时，该模型在不到两小时内发现了 0-day 漏洞，把原本可能需要数周的漏洞研究压缩到了一个下午。团队与受影响的组织合作，以创纪录的速度完成修复，帮助减轻可能波及数万家公司的潜在影响。
In a controlled evaluation of kernel driver vulnerability research, every model we tested found the same core memory-corruption bugs. Carbon for cyber went further, it understood the driver’s constraints, separated the findings that hold up from those that don’t, and uncovered four additional verified vulnerabilities. That is the difference between a finding and an actionable result.
— Chris Corde, Chief Product Officer, Sentinel One
在一次针对内核驱动漏洞研究的受控评估中，我们测试的每个模型都发现了同样的核心内存损坏缺陷。Carbon for cyber 走得更远：它理解该驱动的约束，把站得住脚的发现与站不住脚的发现区分开来，并额外发现了 4 个经过验证的漏洞。这就是「发现」与「可行动结果」之间的差别。
— Chris Corde，Chief Product Officer, Sentinel One
Incorporating deep reasoning models like Carbon for cyber allows defenders to evaluate complex codebases and determine whether vulnerabilities are genuinely exploitable. Our early testing showed a level of precision by Carbon that could allow engineers to move past the noise and focus where human judgment creates the greatest leverage.
— John White, Dir. of Advanced Security Initiatives Group, Cisco
引入像 Carbon for cyber 这样的深度推理模型，能让防御者评估复杂代码库，并判断漏洞是否真正可被利用。我们的早期测试表明，Carbon 达到的精确程度，可以让工程师越过噪声，把精力集中在人类判断能产生最大杠杆的地方。
— John White，Dir. of Advanced Security Initiatives Group, Cisco
Gemini 4 Carbon for cyber is among the highest-scoring models we have ever tested, and across our evaluations it gets there in half the steps of its predecessor. One standout: an exploit that chains multiple hard-to-exploit vulnerabilities into a forged administrator session. Most models we've tested have never completed it. Carbon solved it consistently. This model is a step change for Gemini and a critical tool in the defender's arsenal.
— David Slater, Founder & Chief Architect, Armadin
在我们测试过的模型中，Gemini 4 Carbon for cyber 属于得分最高的之一，而且在我们的评估中，它用前代模型一半的步骤就能达到这一水平。一个突出例子是：把多个难以利用的漏洞串联起来，伪造出管理员会话的利用。我们测试过的大多数模型从未完成过它。Carbon 则能稳定地解决。这个模型对 Gemini 来说是一次阶跃变化，也是防御者武器库中的关键工具。
— David Slater，Founder & Chief Architect, Armadin
Carbon for cyber impressed us right away. Its ability to methodically and systematically catalog large, diverse codebases and generate deep attack chains is best in class. The model's performance even at deep context windows stood out immediately in Chainguard Harness. We'll be integrating it even deeper into our autonomous vulnerability discovery and remediation systems.
— Dan Lorenc, Founder and CEO, Chainguard
Carbon for cyber 一开始就给我们留下了深刻印象。它能够有条理、系统地编目大型、多样的代码库，并生成深入的攻击链，这属于一流水平。即便在很深的上下文窗口下，该模型的表现也在 Chainguard Harness 中立刻脱颖而出。我们将进一步把它更深入地集成到我们的自主漏洞发现与修复系统中。
— Dan Lorenc，Founder and CEO, Chainguard
In early testing, Gemini 4 Carbon for cyber performed well in detecting complex software weaknesses, maintaining the quality of its findings at a highly competitive cost per validated finding compared to other models we tested.
— Carl Windsor, CISO, Fortinet
在早期测试中，Gemini 4 Carbon for cyber 在检测复杂软件弱点方面表现良好，并且在每个经过验证的发现上，能以极具竞争力的成本保持发现质量，与我们测试过的其他模型相比具有优势。
— Carl Windsor，CISO, Fortinet
Stronger scientific and mathematical reasoning更强的科学与数学推理
Gemini 4 Carbon shows strongly improved performance in reasoning across physics, math, and life sciences, synthesizing experimental data, testing hypotheses, and reasoning through multi-step domain problems. For drug toxicity research company Axiom Bio, Carbon is now the leading model in performance for translational science, finding novel mechanisms of toxicity they hadn’t found previously. Axiom is now moving their production scientific analyses and agents to both Carbon and Argon to unlock the creation of safer drugs.
Gemini 4 Carbon 在物理、数学和生命科学中的推理表现有明显提升，能够综合实验数据、检验假说，并推理多步骤的领域问题。对药物毒性研究公司 Axiom Bio 而言，Carbon 现在是转化科学表现领先的模型，发现了他们此前未曾找到的新毒性机制。Axiom 正在把生产环境中的科学分析和 Agent 同时迁移到 Carbon 和 Argon，以推动更安全药物的研发。
Researchers across Google have been using Carbon in Google Antigravity to accelerate their work:
Google 的研究人员一直在 Google Antigravity 中使用 Carbon，以加快工作进度：
• Code verification: Researchers built a new agentic framework for formal verification called AlphaProof Code and put it to work on a major security challenge: mathematically proving the actual code underlying the cryptographic protocol of Signal. Using Carbon alongside other models, our researchers translated security goals to mathematical theorems. AlphaProof Code agents then used Carbon to generate millions of lines of code in Lean, a computer language for writing mathematical proofs, helping them prove key security properties in weeks — a task that historically takes expert teams years. The result indicates Carbon’s potential to unlock scalable formal verification, long considered a gold-standard approach to building secure software.
代码验证：研究人员构建了一个名为 AlphaProof Code 的新智能体形式化验证框架，并把它用在一项重大安全挑战上：从数学上证明 Signal 密码协议底层的实际代码。研究人员把 Carbon 与其他模型一起使用，将安全目标翻译成数学定理。随后，AlphaProof Code Agent 使用 Carbon 生成了数百万行 Lean 代码（Lean 是一种用于编写数学证明的计算机语言），帮助他们在数周内证明了关键安全属性——而这项任务历史上往往需要专家团队花费数年。这一结果表明，Carbon 有潜力打开可扩展的形式化验证，而这一直被视为构建安全软件的金标准方法。
• Enzyme assay development: Provided with only lab notebooks and raw data, Carbon went beyond standard analysis to surface hard-to-interpret data anomalies, which it used to diagnose subtle wet lab experimental artifacts such as incomplete automated mixing in deep wells, leading to substantially improved reproducibility. Moreover Carbon improved methods to deconvolute chromatography peaks and caught trace sample carryover in research instruments, significantly accelerating the development of new assays and tightening the feedback loop between computation and the bench.
酶测定方法开发：在只获得实验记录本和原始数据的情况下，Carbon 超越了常规分析，找出了难以解释的数据异常，并据此诊断出细微的湿实验伪影，例如深孔中自动混合不充分，从而显著提高了可重复性。此外，Carbon 改进了色谱峰解卷积方法，并发现了研究仪器中的微量样品残留，显著加快了新测定方法的开发，也收紧了计算与实验台之间的反馈循环。
• Protein structure prediction: Structure prediction models like AlphaFold 3 use Multiple Sequence Alignments (MSAs) of evolutionary relatives to help infer how amino acid chains fold. As a result they often struggle on understudied, viral, microbial, and eukaryotic proteins that have very few matches in standard databases. Working in a multi-agent setup, our researchers used Gemini 4 Carbon to autonomously discover remote homologs across petabase-scale unannotated sequence archives, environmental and viral sequence catalogs, and 3D structure databases. For example, across 7 low-homolog protein families from the CASP17 blind test set, Carbon increased the number of homologs retrieved by 40-fold on average, greatly increasing the information available to structure prediction models without need for retraining. While broader benchmarking is still underway, addressing this upstream search bottleneck potentially offers a promising path toward modeling hard-to-characterize regions of the proteome.
蛋白质结构预测：像 AlphaFold 3 这样的结构预测模型，会使用进化亲缘序列的多序列比对（MSA）来帮助推断氨基酸链如何折叠。因此，它们往往在研究较少的病毒、微生物和真核蛋白质上遇到困难，因为这些蛋白质在标准数据库中只有极少匹配。在一个多 Agent 设置中，我们的研究人员使用 Gemini 4 Carbon，在拍字节规模的未注释序列档案、环境和病毒序列目录，以及 3D 结构数据库中，自主发现远缘同源物。例如，在 CASP17 盲测集的 7 个低同源蛋白质家族中，Carbon 平均把检索到的同源物数量提高了 40 倍，从而在无需重新训练的情况下，大幅增加了结构预测模型可利用的信息。更广泛的基准测试仍在进行中，但解决这一上游搜索瓶颈，有可能为建模蛋白质组中难以刻画的区域提供一条有前景的路径。
Four weeks ago, we opened applications to give 10,000 scientists access to Argon in Google Antigravity free of charge for six months to accelerate the pace of their research, and we’ve onboarded thousands of scientists already. Today, we’re expanding access to Carbon for current and future participants, and going forward, the program will continue providing access to our latest frontier models as they become available. Applications remain open and onboarding continues on a weekly rolling basis, subject to eligibility, capacity, and selection.
四周前，我们开放了申请，让 10,000 名科学家可以在 Google Antigravity 中免费使用 Argon 六个月，以加快研究进度；目前已经有数千名科学家完成接入。今天，我们向现有和未来的参与者扩大 Carbon 的使用权限；今后，该计划也将继续随着我们最新前沿模型的发布而提供访问。申请仍然开放，接入工作按每周滚动进行，取决于资格、容量和遴选。
Progress in model safety and security模型安全与安保方面的进展
With each new model release, we continue to strengthen our safety and security mitigations, agentic misuse prevention, model alignment, and harm reduction proportionately with model capability.
每一次发布新模型，我们都会继续按模型能力的提升，成比例地加强安全与安保缓解措施、智能体滥用防范、模型对齐，以及危害减少。
Because Gemini 4 Carbon introduces stronger cyber capabilities we build on Argon’s strong safety and security stack to fortify robustness to misuse and prompt injection.
由于 Gemini 4 Carbon 引入了更强的网络能力，我们在 Argon 已有的坚实安全与安保体系之上继续加固，以增强对滥用和提示注入的稳健性。
• Mitigating frontier safety misuse risks: We’ve strengthened our misuse safeguards to help ensure stronger defensive reasoning does not translate into offensive risk. In simulated cyberattack testing environments seeded with ExploitBench and CyScenarioBench, Carbon blocked 99% of attack trajectories from our adaptive red-teaming agent that can iteratively revisit its attack plans after each block. The Carbon safeguards lowered the per-attempt breach rate to just 0.8% (down from 3.1% on Argon). Our safeguards also lowered the automated cyberattack query success rate to 2.17% in Gray Swan’s testing, compared to 9.86% on Argon. And external testing by Trajectory Labs found zero new universal jailbreaks across 132 hours and over 13,000 requests.
缓解前沿安全滥用风险：我们加强了防滥用保障，以帮助确保更强的防御推理不会转化为进攻性风险。在以 ExploitBench 和 CyScenarioBench 为种子的模拟网络攻击测试环境中，Carbon 拦截了来自我们自适应红队 Agent 的 99% 攻击轨迹；该红队 Agent 可以在每次被拦截后迭代地重新制定攻击计划。Carbon 的保障措施把每次尝试的突破率降到了仅 0.8%（Argon 为 3.1%）。我们的保障措施还把 Gray Swan 测试中自动化网络攻击查询的成功率降到了 2.17%，而 Argon 为 9.86%。此外，Trajectory Labs 的外部测试在 132 小时、超过 13,000 次请求中，未发现新的通用越狱。
• Promoting safe and aligned model interactions: When we announced Carbon, we showed significant improvements in alignment as measured by our monitor flag rates across the model’s lifecycle, as well as reductions in data fabrication. New results show that on our 375-task agentic honesty benchmark, it passed 92% of tasks (compared to 70% for Argon) — declining to fill in missing experimental data 98% of the time and flagging methodological flaws 68% of the time.
促进安全且对齐的模型交互：在我们宣布 Carbon 时，就已经展示了对齐方面的显著改进，这体现在模型生命周期中监控标记率的变化，以及数据编造的减少。新结果显示，在我们包含 375 项任务的智能体诚实性基准上，它通过了 92% 的任务（Argon 为 70%）——在 98% 的情况下拒绝补全缺失的实验数据，并在 68% 的情况下标出方法学缺陷。
A chart showing measures of misalignment across a model's lifecycle.
一张展示模型生命周期中未对齐度量的图表。
• Protecting users from indirect prompt injection attacks: Two weeks ago, we shared that Carbon achieved a 0% prompt injection attack success rate on the Gray Swan IPI benchmark — marking the first time that milestone has been reached. At the same time, we know that prompt injection robustness remains an area for continued improvement. Our advanced automated red teaming continues to push the limits here and has shown that models could be further improved against sophisticated attacks in some distributions. It provides the data for us to continuously advance our models’ security to the next frontier.
保护用户免受间接提示注入攻击：两周前我们曾介绍，Carbon 在 Gray Swan IPI 基准上实现了 0% 的提示注入攻击成功率——这是该里程碑首次被达到。与此同时，我们也清楚，提示注入稳健性仍然是需要持续改进的领域。我们先进的自动化红队测试仍在不断推高这里的上限，并表明在某些分布上，模型面对复杂攻击还有进一步改进的空间。这些测试为我们持续把模型安全推进到下一前沿提供了数据。
For more details, read the full Gemini 4 Carbon Safety and Security report.
如需了解更多细节，请阅读完整的 Gemini 4 Carbon 安全与安保报告。
Getting started with Gemini 4 Carbon开始使用 Gemini 4 Carbon
Starting today, Carbon joins Argon in our family of frontier models across our products:
从今天起，Carbon 与 Argon 一起，加入我们各产品中的前沿模型家族：
• For developers: Build with the Gemini API paid tier in AI Studio and Android Studio, and if you’re a Google AI Ultra subscriber, you can try it in Google Antigravity and AI Studio.
面向开发者：通过 AI Studio 和 Android Studio 中的 Gemini API 付费层级来构建应用；如果你是 Google AI Ultra 订阅用户，还可以在 Google Antigravity 和 AI Studio 中试用。
• For enterprises: Try it in Gemini for business via the API or model drop down, Google Antigravity, or CodeMender.
面向企业：可通过 API 或模型下拉菜单，在 Gemini for business、Google Antigravity 或 CodeMender 中试用。
• For everyday users: Try Carbon starting today in the Gemini app and AI Mode in Search if you’re a Google AI Ultra subscriber.
面向日常用户：如果你是 Google AI Ultra 订阅用户，从今天起就可以在 Gemini 应用 和 Search 中的 AI Mode 试用 Carbon。
• For scientists & researchers:Apply for a chance to get access to both Argon and Carbon via Google Antigravity’s highest usage tier free of charge for 6 months (onboarded on a rolling basis), subject to eligibility, capacity, and selection.
面向科学家和研究人员：可以申请，争取免费获得 Argon 和 Carbon 在 Google Antigravity 最高用量层级上为期 6 个月的访问权限（滚动接入），取决于资格、容量和遴选。
• For cybersecurity defenders: Apply to join the Fairwind Program to get access to Carbon for cyber on a rolling basis.
面向网络安全防御者：可申请加入 Fairwind 计划，以滚动方式获得 Carbon for cyber 的访问权限。
See the model card to learn more about Gemini 4 Carbon and to start building with it, explore our updated Gemini 4 developer guide and documentation for best practices on migrating workflows to the model.
可查看模型卡，了解更多关于 Gemini 4 Carbon 的信息；要开始用它构建应用，也可参阅我们更新后的 Gemini 4 开发者指南和文档，了解把工作流迁移到该模型的最佳实践。
Footnote脚注
1. After the introductory period expires on November 30, 2026, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply. Applies to the Carbon model intended for public access only.
1. 导入期于 2026 年 11 月 30 日结束后，将适用每 100 万输入 token 4 美元、每 100 万输出 token 20 美元的价格。该价格仅适用于面向公众提供的 Carbon 模型。
About the Author关于作者
Koray Kavukcuoglu
SVP, Google DeepMind and Chief AI Architect, Google
Koray Kavukcuoglu
Google DeepMind 高级副总裁、Google 首席 AI 架构师
Koray is one of the world's foremost experts in artificial intelligence. As the SVP of Google DeepMind, he leads the development of Google DeepMind’s state-of-the art generative AI models and their integration into Google’s products at scale. Prior to his current role, he was VP of Research at Google DeepMind, where he founded the deep learning team and his team pioneered major research breakthroughs such as DQN (Deep Q-Network) and WaveNet.
Koray 是世界上最重要的人工智能专家之一。作为 Google DeepMind 高级副总裁，他领导 Google DeepMind 最先进生成式 AI 模型的开发，以及这些模型大规模融入 Google 产品的工作。在担任现职之前，他曾是 Google DeepMind 的研究副总裁，在那里创办了深度学习团队；该团队开创了 DQN（Deep Q-Network）和 WaveNet 等重大研究突破。
Read more from Koray Kavukcuoglu
阅读 Koray Kavukcuoglu 的更多文章
来源：Google The Keyword · Gemini 4 Carbon is now available — https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/