[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-gemini-4-carbon-cyber-defense-frontier-model":3,"topics-all":41,"news-related-4f733d5a-b430-4c93-8e55-aa061d59557c":60},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"4f733d5a-b430-4c93-8e55-aa061d59557c","Gemini 4 Carbon:谷歌最强模型,首发只给网安","谷歌发布 Gemini 4 Carbon,官方称最强模型,主打网安与长周期任务。它在 Artificial Analysis 指数以 59 分登顶,SEC-bench Pro 88.2%、ExploitGym 54.8%、CWE-bench 71.0% 领先,但跑分由谷歌自测。首发只给网安防守者,开发者暂无 API。","Google 上个月底发布的 Gemini 4 Argon 之后,10 月 28 日又来了一个 Gemini 4。Google DeepMind 的 SVP、首席 AI 架构师 Koray Kavukcuoglu 在[官方公告](https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Fmodels-and-research\u002Fgemini-models\u002Fintroducing-gemini-4-carbon\u002F)里说,这是\"我们最强的模型\",主打方向不是通用对话,而是网络安全防御和长周期工程任务。名字 Carbon 取的是碳——碳纤维复合材料那一层,碳元素最硬的一种形态。\n\n有意思的是,这份公告同时是成绩单和限售通知。而成绩单里最耐看的部分,恰恰是它由谁来打分。\n\n## 登顶是真的,但榜单不是自己出的\n\n最硬的一个数是 Artificial Analysis Intelligence Index v5.0 里的 59 分,排名第一。谷歌的公告这么说,第三方评测机构 Artificial Analysis 的榜单页也这么说——那个页面上 Carbon 后面跟着 Claude Opus 5.5 的 58 分,再后面是 GPT-6 Astra 的 57 分。MarkTechPost 汇总的对比表同样记录 Carbon 59、Opus 5.5 58、Astra 57。\n\n这一项没有自说自话,至少在综合能力上。\n\n## 网安三项领先,但跑分方法是自家人\n\nCarbon 的叙事是一套完整的攻防闭环:找漏洞、验证可利用性、修根因。三个数字分别是 SEC-bench Pro 88.2%、ExploitGym 54.8%、CWE-bench 71.0%,谷歌都称为 SOTA。\n\n对比数据同样来自谷歌公布的那张表:ExploitGym 上 GPT-6 Astra 是 42.4%、Claude Opus 5.5 是 34.5%、Argon 是 42.0%;SEC-bench Pro 上 Astra 85.4%、Argon 83.4%。\n\n但这里必须打个折。MarkTechPost 明确写道,SEC-bench Pro 和 ExploitGym 的分数是谷歌用自家的 Antigravity 测试框架自己算的,ExploitGym 运行时缓解措施是关闭的,竞品的分数大多由各家自报。这是一份官方自测表,不是第三方复现,目前也没有独立复现。\n\n谷歌自己倒是给了两个旁证:Chrome 团队称 Carbon 生成高质量补丁的表现比\"另一家实验室的大型前沿模型\"好 10 倍以上(注意,它没点名);云安全平台 Wiz 称 Carbon 在其 Web 渗透测试基准上领先。\n\n## 长任务是主场,通用操作台不是\n\nCarbon 面向的是能连续跑几天的 agentic 工作流:跨大型代码库定位并修补隐蔽缺陷,跨底层系统做根因分析。\n\n数字上差距最悬殊的是 GraphWalks,96.2%,GPT-6 Astra 是 71.8%。Terminal-Bench 4.0 上 Carbon 65.2%,Opus 5.5 是 59.6%。\n\n但它不是全面碾压。OSWorld-2.1 上 Carbon 81.5%,落后 Claude Opus 5.5 的 87.2%;FrontierSWE 上 Carbon 61.2%,GPT-6 Astra 以 65.5% 领先。按 MarkTechPost 的统计,Carbon 在 17 项公开基准里领先 15 项。\n\n参数量和上下文窗口,两个来源都显示未披露。没有开源权重,没有自托管,没有量化版本,只有谷歌托管这一条路。\n\n## 安全数字这次被摆到了台面上\n\n这一代的安全指标有点反常地抢眼:Terminal-Bench 4.0 上成功作弊率压到 0.0%,尝试作弊的比例只有 0.4%;Gray Swan 间接提示注入基准的攻击成功率 0.0%(Argon 是 0.7%);内部真实编码环境里高危对齐告警下降近 80%;无依据的断言比 Argon 减少约一半;过度介入下降 26%,服务不足下降 29%。\n\n谷歌的说法是:与其在证据不足时含糊其辞,不如把方法论上的保留直接写进交付物,或者干脆拒绝执行并列出还缺什么。\n\n同样要打折扣——这些是内部或自跑数据,谷歌自己承认尚未有独立复现。\n\n## 价格不贵,通道很窄\n\n定价与 Argon 的引入价一致:每百万输入 token 2 美元,输出 10 美元,缓存输入按输入价打 95% 折;引入期结束后分别涨到 4 美元和 20 美元。\n\n真正的限制在发布节奏。本周先给一部分网络防御者,之后通过 Fairwind Program 逐步开放给更多可信伙伴——这个项目优先考虑政府、关键基础设施运营方和核心技术平台。开发者、企业和普通用户什么时候能用,谷歌没给日期。谷歌同时表示正在参与美国政府的自愿预发布模型测试流程。\n\n还有个容易被忽略的细节:现在能在 Gemini API 上调用的 Gemini 4 模型还是 Argon,不是 Carbon。\n\n## 所以呢\n\n把 Carbon 单独拎出来看,它更像一次能力定投:谷歌押的不是\"更大更强的通用模型\",而是把模型塞进安全防御这个高价值、且监管友好的垂直场景里,用限售的方式先跑一段。\n\n值得注意的是这一代把对齐数字和安全数字放在同一张成绩单上一起发布。对一个能力正在快速逼近攻击用途的模型来说,这种透明可能是策略的一部分——先让防守方拿到,划出使用边界,再谈开放。\n\n不过这也留下一个问题:当最强的一档能力只对一部分人开放时,评测榜单上排第一的模型,和所有人手里的模型之间,会隔出多大的距离?如果这层距离持续拉大,基准排名本身也会慢慢失去参照意义。\n\n两个月内连发两个 Gemini 4,中间还夹着一个只给网安群体的型号。谷歌这条产品线的节奏越来越像软件版本迭代,而不是模型发布——那么下一次 Carbon 级别的更新,还打算等多久?\n","https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Fmodels-and-research\u002Fgemini-models\u002Fintroducing-gemini-4-carbon\u002F","d884df39-706a-45db-b7ed-371f12e54f1f",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"a9524a82-a7c5-4daa-bb4b-a7ee77bb0b94","gemini",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"8cf7490f-2449-4ba7-be19-61befa0d92b4","google",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":25,"name":26,"slug":26,"description":14,"color":14},"02bfdc0f-5e23-4fbb-b31a-7cc03cdba1ee","security",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"d738ac3d-f2bc-48b0-bfac-a978d402b913","en","Gemini 4 Carbon:Google's Top Model, Cyber Defenders First","Gemini 4 Carbon tops the Artificial Analysis Index at 59 and leads cyber benchmarks, self-run by Google. Cyber defenders get it first; developers get no API.","Two months after Gemini 4 Argon, Google has shipped another Gemini 4. On October 28, Koray Kavukcuoglu, Google's SVP of DeepMind and Chief AI Architect, announced in the [official announcement](https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Fmodels-and-research\u002Fgemini-models\u002Fintroducing-gemini-4-carbon\u002F) that Gemini 4 Carbon is \"our strongest model,\" aimed not at general chat but at cybersecurity defense and long-horizon engineering tasks. The name Carbon points at carbon fiber composites, the hardest structural form of the element.\n\nWhat makes this announcement unusual is that it works as two documents at once: a scorecard and a limited-availability notice. And the most interesting part of the scorecard is who did the scoring.\n\n## The top spot is real, and it is not self-awarded\n\nThe hardest single number is a score of 59 on version 5.0 of the Artificial Analysis Intelligence Index, which puts Carbon in first place. Google's announcement says so, and so does the third-party evaluator's own leaderboard page, where Carbon is followed by Claude Opus 5.5 at 58 and GPT-6 Astra at 57. MarkTechPost's comparison table records the same three figures: Carbon 59, Opus 5.5 58, Astra 57.\n\nOn this one dimension, at least, nobody is marking their own homework.\n\n## Three cybersecurity wins, measured by the family\n\nCarbon's story is a complete defense lifecycle: find the vulnerability, validate whether it is exploitable, then fix the root cause. The three numbers are 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench, all of which Google calls state of the art.\n\nThe comparison figures come from that same published table: on ExploitGym, GPT-6 Astra scores 42.4%, Claude Opus 5.5 scores 34.5%, and Argon scores 42.0%; on SEC-bench Pro, Astra scores 85.4% and Argon 83.4%.\n\nBut there is a discount to apply here. MarkTechPost states plainly that the SEC-bench Pro and ExploitGym scores were computed by Google using its own Antigravity test harness, that ExploitGym ran with mitigations switched off, and that most rival numbers are self-reported by each provider. This is an official self-run table, not third-party replication, and no independent replication exists yet.\n\nGoogle does offer two corroborating notes of its own: the Chrome team reports that Carbon produced high-quality patches more than 10x better than a large frontier model from \"another lab\" (notice it declines to name them), and cloud security platform Wiz says Carbon leads its web penetration testing benchmarks.\n\n## Long tasks are the home turf, general computer use is not\n\nCarbon targets agentic workflows that can run for days at a stretch: root-causing and patching subtle bugs across large codebases, working through low-level systems.\n\nThe widest gap shows up on GraphWalks, where Carbon scores 96.2% against GPT-6 Astra's 71.8%. On Terminal-Bench 4.0, Carbon reaches 65.2% versus Opus 5.5 at 59.6%.\n\nIt is not a clean sweep everywhere. On OSWorld-2.1, Carbon's 81.5% trails Claude Opus 5.5 at 87.2%, and on FrontierSWE, Carbon's 61.2% is beaten by GPT-6 Astra at 65.5%. By MarkTechPost's count, Carbon leads 15 of 17 published benchmarks.\n\nParameter count and context window both remain undisclosed. There are no open weights, no self-hosting path, and no quantized variants, only Google's hosted service.\n\n## The safety numbers get put on the table this time\n\nThis generation's safety figures are unusually prominent: successful cheating on Terminal-Bench 4.0 drops to 0.0%, with attempts to cheat falling to 0.4%; attack success on Gray Swan's indirect prompt injection benchmark reaches 0.0%, against 0.7% for Argon; high-severity misalignment alerts in live internal coding environments fall by almost 80%; unsupported claims drop roughly 2x versus Argon; overreach falls 26% and under-service falls 29%.\n\nGoogle's framing is that rather than hedging or fabricating around missing evidence, the model should put methodological caveats directly into the deliverable, or withhold the blocked action with a clear list of what is still required.\n\nThe same discount applies. These are internal or self-run figures, and Google itself notes that independent replication is not yet available.\n\n## The price is cheap, the channel is narrow\n\nPricing matches Argon's introductory rate: $2 per million input tokens and $10 per million output tokens, with cached input priced at 95% off the input rate; after the introductory period, those rise to $4 and $20.\n\nThe real constraint is the rollout. A subset of cyber defenders gets access this week, with more trusted partners phased in through the Fairwind Program, which prioritizes governments, critical infrastructure operators, and core technology platforms. Google gave no date for developers, businesses, and consumers. The company also says it is participating in the U.S. government's voluntary pre-release model testing process.\n\nOne easily missed detail: the Gemini 4 model you can call on the Gemini API today is still Argon, not Carbon.\n\n## So what\n\nTaken on its own, Carbon looks like a targeted bet on capability rather than a general push for a bigger model. Google is not betting on \"larger and stronger\" but on slotting the model into cybersecurity defense, a high-value and regulation-friendly vertical, and running it in limited release first.\n\nWhat is notable is that this generation publishes the alignment and safety numbers on the same scorecard as the capability numbers. For a model whose abilities are rapidly approaching offensive use, that transparency may itself be part of the strategy: hand it to defenders first, draw the usage boundary, then talk about opening it up.\n\nThat leaves a question. When the strongest tier of capability is open only to part of the population, how wide does the gap grow between the model at the top of the leaderboard and the model in everyone's hands? If that distance keeps widening, benchmark rankings slowly stop being a useful reference point.\n\nTwo Gemini 4 models in two months, with a cyber-only variant slotted in between. Google's release cadence on this line increasingly looks like software version bumps rather than model launches, which raises the obvious question: how long until the next Carbon-class update?\n","gemini-4-carbon-cyber-defense-frontier-model","2026-10-29T03:05:00Z","2026-10-29T03:12:23.184516Z","2026-10-29T03:12:23.184524Z",true,"agent",179,[42,51],{"slug":43,"tag_slug":43,"title_zh":44,"title_en":45,"intro_zh":46,"intro_en":47,"id":48,"is_active":38,"created_at":49,"modified_at":50},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":52,"tag_slug":52,"title_zh":53,"title_en":54,"intro_zh":55,"intro_en":56,"id":57,"is_active":38,"created_at":58,"modified_at":59},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":61},[62,67,72,77,82,87],{"id":63,"title":64,"news_slug":65,"published_at":66},"260aa193-a883-4db9-9e05-7bf3c6f14f2f","Gemini 4 Argon 发布:输出上限 100 万 token,首批给安全防御者","gemini-4-argon-release","2026-10-01T13:11:08+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"039ff515-68e7-4f11-866a-1da97e26eb45","Gemini 3.8 Live 拿下 S2S 实时语音榜第一","gemini-3-8-live-voice-s2s-number-one","2026-09-15T17:00:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"e0a484e4-41e0-4f91-9b9b-a196bbdcf3ba","Gemini 3.8 Flash 双发:同价升级 + Cyber 走可信项目 Fairwind","gemini-3-8-flash-cyber-fairwind-launch","2026-09-03T03:00:00+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"16856034-439d-4915-aed4-80b42ae09c68","Gemini 3.7 Flash：FrontierCode 43.6%，价格腰斩","gemini-3-7-flash-coding-agent-fast","2026-08-13T09:00:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"4d1b266e-9e0b-4dd7-85b3-adb5a3c94c7e","Gemini Omni 1.2 Flash:0.07美元每秒","gemini-omni-1-2-flash-video-pricing-drafts","2026-11-05T10:00:00+00:00",{"id":88,"title":89,"news_slug":90,"published_at":91},"f093935e-e23c-4766-88d6-9302a370be2a","Gemini 套餐洗牌:免费版限 Flash-Lite","gemini-deep-think-ai-pro-tier-restructure","2026-10-09T00:00:00+00:00"]