Google 自家最强模型,2026 年 11 月 10 日全面开放:主打长程 Agent 任务、复杂编程、网络安全与科学推理。API 与 Argon 同价($2/$10 至 11 月 30 日),但思考 token 更多、速度更慢;Artificial Analysis 指数 60,曾短暂登顶,两天后被 Claude Fable 5.2 超过。本页汇总官方规格、上手路线、价格与中英文首批实测,并明确标注哪些说法尚未证实。
先把能确认的说清楚:这是一个正式版(GA),不是限时预览;但它的「最强」头衔只维持了两天。
gemini-4-carbon(API 稳定版标识)从宣布到 GA 只隔了 13 天;登顶 AA 指数只维持了约两天。
Google 在 blog.google 宣布 Gemini 4 Carbon;同日仅向 Fairwind Program 的部分网络安全防御方开放,专项网络安全配置至今仍需申请。公告称 AA 指数 v5.0 得 59 分、排名第一。
LMArena Text 预览分约 1528,暂列第一(约 4,403 票)。
Artificial Analysis 指数更新至 v5.0.1:Carbon(High)60 分,登顶;Opus 5.5 为 59、Astra 为 57。
Claude Fable 5.2 发布,AA 指数 65,超过 Carbon 夺回第一;FrontierSWE v2 亦以 69.9% 领先 Carbon 的 61.2%。
全面开放(GA):付费 API、AI Studio、Gemini for business、Ultra 订阅(Gemini App、AI Mode、Antigravity);模型卡发布;Antigravity 推出 /swarm 多 Agent 预览;Cursor 当天接入。
首批实测潮:Quantslant、虎嗅、鱼皮、SeedRouter、eesel 等中英文测评陆续发布(见第 06 节)。
API 优惠价截止($2/$10)。
API 价格翻倍至 $4 输入 / $20 输出(每百万 tokens)。
以下为官方已公布口径(模型页与模型卡,2026 年 11 月 11 日核查);未公布的项直接写明,不做推测填充。
开发者最快的是付费 API;普通用户只有 Ultra 订阅一条路——Pro 订阅没有 Carbon。
开发者主路线。付费层 API Key 即可调用,模型 ID gemini-4-carbon;从 Argon 迁移只需改模型名,请求结构不变。免费层没有 Carbon,需要绑定账单账户并预付。
Google AI Ultra 订阅用户在 Gemini App、Search AI Mode、Antigravity 中使用,模型选择器提供 Low / Medium / High 三档(默认 High)。Free、Plus、Pro 订阅均无。
企业路线:API 或下拉菜单调用,Antigravity 与 CodeMender 中可用。
Google Antigravity 中的多 Agent 编排预览:一个提示词派出一队协调 Agent,可把 Carbon 指派给架构、协调、审计角色,搭配更快更轻的模型与对抗验证器。
OpenRouter 上为 google/gemini-4-carbon,与官方同价 $2/$10,:batch 变体 $1/$5;models.dev 目录有 16 家提供商列出 Carbon。Cursor 在 11 月 10 日当天接入。
科学家:10,000 名科学家项目入选者免费 6 个月在 Antigravity 最高用量档用 Argon + Carbon,申请开放、每周滚动入选。网络安全:专项配置仍通过 Fairwind Program 申请。
单价与 Argon 完全相同,但账单未必相同——Carbon 每个任务花的思考 token 更多。价格为 2026 年 11 月 11–12 日核查口径,结账页价格为准。
| API 档位(每百万 tokens) | 至 2026-11-30 | 2026-12-01 起 |
|---|---|---|
| Standard 输入 | $2.00 | $4.00 |
| Standard 输出(含思考 tokens) | $10.00 | $20.00 |
| 缓存输入(隐式缓存,95% 折扣) | $0.10 | $0.20 |
| Batch API(输入 / 输出) | $1 / $5 | $2 / $10 |
| Flex(输入 / 输出) | $1 / $5 | $2 / $10 |
| Priority(输入 / 输出) | $3.60 / $18 | $7.20 / $36 |
缓存为隐式缓存:提示不少于 4,096 tokens 时自动触发。Batch / Flex / Priority 价格引自定价页梳理(agentdeals.dev,11 月 12 日)。
美国定价 $99.99/月(Ultra 5x,Pro 用量 5 倍)/ $199.99/月(Ultra 20x)。Ultra 在 150+ 国家销售;Carbon 覆盖 Ultra 用户的「部分国家」,Google 未列举具体名单。Pro、Plus、免费订阅均无 Carbon——这一分层在 11 月 10–11 日的 Instagram / Facebook 评论中引发付费 Pro 用户不满,不少人拿 Claude Pro 提供 Fable 5.2 作对比。
Artificial Analysis 按标准价测得:Carbon(High)每任务 $7.76,约比 Claude Opus 5.5(max,$10.04)低 23%,约 1.2 倍于 GPT-6 Astra(max,$6.25);Fable 5.2(High)为 $9.08,只比 Carbon 贵 1 美元多。分档看:Medium 每任务 $6.59 反而比 Low($7.19)便宜——弱档位更容易绕圈子、耗更多 token。简单任务上 Carbon 明显更贵:同一组 11 个小任务,Carbon 默认档成本是 Argon 的 3.05 倍;但在编码 Agent 任务上,Carbon 为 $7.74/任务,反而低于 Argon 的 $11.73。
以下全部是 11 月 10–12 日的首批上手实测:样本小、方法不一,请按「个案」看待。每条都给出结论徽章;点「阅读本站全文」进入本站的深度改写文章页,文末附原始出处。
来源:Quantslant,2026-11-11(vs Argon、Fable 5.2、GPT-6.1 Sol,同设置、单元测试评分)
23 项实用任务 Carbon 99.5%($0.23,264 秒),与 Argon 持平但慢一倍;Agent、电子表格(34/34)、安全审查(8/8)、SOC 日志(7/7)全部满分。但 6 项编码任务、76 个单元测试中 Carbon 仅 63/76,三位对手全部 76/76。另有思考烧光 Bug:正则任务 medium 档用掉全部 16,000 tokens 推理,未返回代码。
阅读本站全文 → (本站原创整理)
来源:SeedRouter / blog.laozhang.ai,2026-11-11
同提示词测试:两者答案都对,但 Carbon 计费输出 token 多约 44%(小样本);11 个小任务上 Carbon 成本是 Argon 的 3.05 倍、耗时 2.75 倍。反过来:1,500 行日志精确计数,Argon 22 秒答错,Carbon 214 秒答对;编码 Agent 任务上 Carbon 每任务 $7.74,反而低于 Argon 的 $11.73。Google 的建议:交互式编程用 Argon,Argon 答错时换 Carbon。
阅读本站全文 → (本站原创整理)
来源:eesel.ai / Inferera,2026-11-11/12
AA 三档数据:Low 54 分、$7.19/任务、编造率 20%;Medium 58 分、$6.59/任务、编造率 4.2%(最便宜也最诚实);High 60 分、$7.76/任务、编造率 11%。代价是时间:首个词约等 1 分钟,high 档每基准任务约 26 分钟。API 侧:thinking_budget、minimal 档、candidate_count>1 均返回 400;低输出上限会静默截断思考、返回空文本且照样收费;建议 HTTP 超时 15–20 分钟。
阅读本站全文 → (本站原创整理)
来源:Google 官方 10 月 28 日分数表;Artificial Analysis;LMArena,2026-11-07/12
Google 官方 17 项表:Carbon 15 项第一,只输 FrontierSWE 2(61.2%,低于 Astra 的 65.5%)与 OSWorld-2.1(81.5%,低于 Opus 5.5 的 87.2%)。但多项分数由 Google 自行计算,安全测试在无网、防护关闭环境下进行,且对比的是上一代 Fable 5.1。独立榜单:AA 指数 60(11 月 7 日登顶,11 月 9 日被 Fable 5.2 的 65 超过)、LMArena 1526 预览分第一、WebDev Arena 第 4。
阅读本站全文 → (本站原创整理)
来源:虎嗅,2026-11-11(Antigravity,High 档,全部一次成型)
9 个任务没有一项彻底失败:户型图变可漫游 3D 楼栋(约 1 小时)、iPhone Duo 交互 3D 展品、手写布料物理丝绸卷轴、逐帧复刻 20 秒宣传片(37 个硬切与原片一致)、8D 重庆夜景等。短板是精细度:家具简单、材质重复、屏幕内容超出圆角机身;15 秒艺术史第一版还「作弊」调用了图像生成。作者认为 Google 不再宣称「最强」,转而强调任务成本不到第二名一半、价格约 Fable 5.2 的四分之一。
阅读本站全文 → (本站原创整理)
来源:腾讯云开发者社区,2026-11-11(Cursor,3 个并行子 Agent,High 档,复用 Fable 5.2 同题)
「上古卷轴」被服务端错误 3 次打断,耗时 2 小时 16 分钟(Fable 5.2 为 86 分钟、零中断);重庆 8D 城市生成器 34 分钟交卷但只有 184 栋方盒子楼(Fable 5.2 为 1,230 栋);溶洞项目第一轮近 2 小时后作废重来,大黄狗是圆柱拼的、镜头会穿进岩壁。3 个项目共花 $21.30,远低于 Fable 5.2 的 500 多块;作者结论:「不要完全相信跑分」。
阅读本站全文 → (本站原创整理)
以下 6 篇均为本站原创改写与再整理的深度文章——重新组织结构、补充结论与选型建议,事实与数据与原始测评核对一致。点击任意一篇在本站内阅读,文末附原始出处。
Carbon 与 Qwen4-Max 相反:分数很多,问题是谁测的、跟谁比。下面先给 Google 官方 17 项表(2026-10-28),再给独立榜单,最后说 caveat。
| 测试(Google 官方表) | Carbon | Argon | GPT-6 Astra | Opus 5.5 | Fable 5.1 |
|---|---|---|---|---|---|
| AA Index v5.0 | 59 | 50 | 57 | 58 | 49 |
| Vals Index v2.2 | 70.4% | 69.2% | 65.0% | 70.2% | — |
| SWE-sweep | 9.6% | 4.3% | 9.5% | 3.6% | 2.9% |
| FrontierSWE 2 | 61.2% | 55.0% | 65.5% | 62.3% | 56.3% |
| Terminal-Bench 4.0 (AA) | 65.2% | 57.1% | 59.1% | 59.6% | 52.0% |
| Agents' Last Exam v1.1 | 77.4% | 69.0% | 73.1% | 71.6% | — |
| OSWorld-2.1(离线) | 81.5% | 74.7% | 80.9% | 87.2% | 76.8% |
| AutomationBench (Zapier) | 53.9% | 51.3% | 41.4% | 42.5% | 31.4% |
| Vending-Bench 2(美元余额) | 16,484 | 13,718 | 15,515 | 9,235 | 5,422 |
| Chartography | 72.5% | 71.6% | 71.0% | 66.3% | 46.2% |
| Blueprint-Bench 2 | 58.6% | 54.4% | 49.7% | 51.2% | 41.9% |
| GraphWalks 256K–1M | 96.2% | 84.2% | 71.8% | 66.8% | 65.0% |
| RiemannBench | 80.0% | 76.0% | 72.0% | 69.6% | 65.6% |
| Terminal-Bench-Science 0.1 (AA) | 64.8% | 52.9% | 63.3% | 59.0% | 43.3% |
| CWE-bench | 71.0% | 68.0% | 68.0% | 67.0% | 58.0% |
| ExploitGym | 54.8% | 42.0% | 42.4% | 34.5% | — |
| SEC-bench Pro | 88.2% | 83.4% | 85.4% | — | — |
Carbon 15/17 第一。官方表对比的是上一代 Claude Fable 5.1,而非 11 月 9 日发布的 Fable 5.2。
| 测量 | 来源 | Carbon 成绩 | 对照 |
|---|---|---|---|
| AA 智能指数 v5.0.1 | Artificial Analysis,11 月 7 日 | 60.0(High),曾登顶 | Fable 5.2 65(11 月 9 日起居首);Opus 5.5 59;Astra 57;Argon 50 |
| AA 指数分档 | Artificial Analysis / eesel,11 月 11 日 | Low 54 / Medium 58.3 / High 60 | Medium 每任务 $6.59 最便宜,编造率 4.2% 最低 |
| AA-Omniscience v1.1 | Artificial Analysis,约 10 月 31 日 | 净分 54,第 1 | Opus 5.5 48;准确率 58%(Argon 48%);幻觉率 11%(Argon 4%) |
| LMArena Text | LMArena,11 月 6–7 日 | 1526 预览分,第 1(约 4,483 票) | Argon(High)1523;Muse Wonder 1513;Opus 5.5 1505 |
| WebDev Arena | LMArena | 1768 elo,第 4 | — |
| CursorBench 4.0 | Cursor | 第 1(high 档 60.4%) | — |
| Coding Agent Index | Artificial Analysis | 59 | Claude Code + Opus 5.5 为 62,Carbon 落后 3 分 |
| MKA(多模态知识 Agent) | Google × AA 合作新基准 | 第 2 | 平均任务成本不到第一名的 1/6 |
| FrontierSWE v2 | Snowtraces,11 月 11 日 | 61.2% | Fable 5.2 69.9% |
完整盘点见本站基准测试专文 →
| 维度 | Gemini 4 Carbon | Gemini 4 Argon |
|---|---|---|
| 状态 | GA(2026-11-10) | GA(更早发布) |
| API 价格 | $2 / $10 至 11 月 30 日,之后 $4 / $20 | 与 Carbon 完全相同 |
| 上下文 | 1M | 1M |
| AA 指数 | 60(High)/ 58(Medium)/ 54(Low) | 50(High) |
| 速度 | 慢:每任务约 26–27 分钟,首 token 约 68 秒 | 快得多:同组任务约一半时间 |
| 订阅 | 仅 Ultra | Pro 及以上可用 |
| 适合 | 最难的任务:长程 Agent、疑难调试、科学推理、安全审查 | 大多数编程与知识工作、交互式编程、快速循环 |
Google 自己的建议:Argon 仍是大多数软件工程与知识工作的正确选择;Carbon 留给最难的任务。
| 模型 | API 价格 $/百万(入/出) | AA 指数 | 上下文 |
|---|---|---|---|
| Claude Fable 5.2 | 8 / 40 | 65(max) | 1M |
| Gemini 4 Carbon | 2 / 10(至 11 月 30 日) | 60(high) | 1M |
| Claude Opus 5.5 | — | 59(max) | — |
| GPT-6 Astra | 10 / 50 | 57(max) | 1.05M |
| GPT-6.1 Sol | 2 / 10 | 55(max) | 1.05M |
| Qwen3.8 Max | 2 / 6 | 45(09-27) | 1M |
| Qwen4-Max-Preview | 仅 Token Plan 订阅 | 未收录 | 1M |
Carbon 没有出现大规模谣言潮,但以下说法正在流传或容易混淆,需要逐条澄清。
都是真的,只是版本不同:59 是 AA 指数 v5.0(10 月 28 日公告时),60 是 v5.0.1(11 月 7 日更新后)。引用时请注明版本。
曾是,但只维持约两天(11 月 7–9 日)。11 月 9 日 Claude Fable 5.2 发布后以 65 分居首,截至 11 月 12 日 Carbon 60 分居其后。另注意:不同统计口径(按模型 / 按配置)给出的 Carbon 排名在 2–4 位之间,本页不写单一排名数字。
AA 指数(65 对 60)、FrontierSWE v2(69.9% 对 61.2%)均为 Fable 5.2 领先。Google 官方对比表用的是上一代 Fable 5.1——这张表发布时 Fable 5.2 还未发布,但现在拿它论证「Carbon 最强」已经过时。
Google 自述 Carbon 基于 Argon 的预训练基础与架构,「主要差异是每轮花多少思考 token」,且单价完全相同。但这不等于「同一个模型」:额外后训练带来了实质改进——作弊尝试 3.6%→0.4%(成功 2.4%→0%)、Agent 诚实基准 70%→92%、AA 指数 50→60。
定价页明确 Carbon 不在免费层;Google AI Pro 订阅同样没有,Pro 仍为 Argon。消费者侧只有 Ultra 订阅能用。
2026 年 10 月 28 日宣布,11 月 10 日全面开放(GA)。10 月 28 日至 11 月 10 日之间仅开放给 Fairwind Program 的部分网络安全防御方。时间线见本页第 02 节。
不免费。API 没有免费层;消费者侧只有 Google AI Ultra 订阅($99.99 / $199.99 每月,美国)能用,Pro、Plus、免费订阅均无。科学家计划入选者可免费使用 6 个月。
至 11 月 30 日:输入 $2 / 输出 $10(每百万 tokens,含思考 tokens);12 月 1 日起翻倍至 $4 / $20。缓存输入 95% 折扣($0.10 → $0.20)。Batch 与 Flex 为半价。价格见本页第 05 节。
AA 指数 60 对 50,安全与诚实指标明显更好;但日常任务上它做不了 Argon 做不了的事,还更慢更贵。Google 的建议:交互式编程与大多数知识工作用 Argon;Argon 答错或漏东西时换 Carbon。详见本站Carbon vs Argon 对比。
按分数不强:AA 指数 65 对 60、FrontierSWE v2 69.9% 对 61.2%,均为 Fable 5.2 领先;前端实测中 Fable 5.2 的审美与稳定性也更好。但 Carbon 单价只有 Fable 5.2 的四分之一,任务成本 $7.76 对 $9.08,差距只有 1 美元多。
默认 medium 是最划算的档:AA 58 分、每任务 $6.59、编造率 4.2%(三档中最低);High 60 分但每任务 $7.76、编造率 11%;Low 54 分、每任务 $7.19、编造率 20%。详见本站思考档位与避坑指南。
输出上限覆盖思考 token:上限设太低时,模型把 token 全花在思考上、还没写答案就触顶,返回空文本、状态 incomplete,费用照收。这是已确认的行为(Quantslant 实测、Google 文档均有说明),不是偶发故障。解决办法:把 max_output_tokens 设到至少 32,768,用思考档位而不是输出上限控成本。
1,048,576 tokens,输入、思考与输出共享。最大输出 65,536 tokens(流式)或 32,768(非流式)。
不是开源模型,没有发布权重,HuggingFace 上没有官方模型。只能通过 Google API 与订阅服务使用。
GA 公告材料未说明 Vertex AI 的可用性,截至 11 月 13 日没有确认消息。请以 Google Cloud 官方文档为准。
Google 官方博客(blog.google)、Gemini API 定价页与模型页、Artificial Analysis 模型页,以及本页下方的来源列表。
本页事实均来自以下公开来源(核查于 2026 年 11 月 12–13 日)。
Google 自家最強模型,2026 年 11 月 10 日全面開放:主打長程 Agent 任務、複雜編程、網絡安全與科學推理。API 與 Argon 同價($2/$10 至 11 月 30 日),但思考 token 更多、速度更慢;Artificial Analysis 指數 60,曾短暫登頂,兩天後被 Claude Fable 5.2 超過。本頁彙總官方規格、上手路線、價格與中英文首批實測,並明確標註哪些說法尚未證實。
先把能確認的說清楚:這是一個正式版(GA),不是限時預覽;但它的「最強」頭銜只維持了兩天。
gemini-4-carbon(API 穩定版標識)從宣佈到 GA 只隔了 13 天;登頂 AA 指數只維持了約兩天。
Google 在 blog.google 宣佈 Gemini 4 Carbon;同日僅向 Fairwind Program 的部分網絡安全防禦方開放,專項網絡安全配置至今仍需申請。公告稱 AA 指數 v5.0 得 59 分、排名第一。
LMArena Text 預覽分約 1528,暫列第一(約 4,403 票)。
Artificial Analysis 指數更新至 v5.0.1:Carbon(High)60 分,登頂;Opus 5.5 爲 59、Astra 爲 57。
Claude Fable 5.2 發佈,AA 指數 65,超過 Carbon 奪回第一;FrontierSWE v2 亦以 69.9% 領先 Carbon 的 61.2%。
全面開放(GA):付費 API、AI Studio、Gemini for business、Ultra 訂閱(Gemini App、AI Mode、Antigravity);模型卡發佈;Antigravity 推出 /swarm 多 Agent 預覽;Cursor 當天接入。
首批實測潮:Quantslant、虎嗅、魚皮、SeedRouter、eesel 等中英文測評陸續發佈(見第 06 節)。
API 優惠價截止($2/$10)。
API 價格翻倍至 $4 輸入 / $20 輸出(每百萬 tokens)。
以下爲官方已公佈口徑(模型頁與模型卡,2026 年 11 月 11 日覈查);未公佈的項直接寫明,不做推測填充。
開發者最快的是付費 API;普通用戶只有 Ultra 訂閱一條路——Pro 訂閱沒有 Carbon。
開發者主路線。付費層 API Key 即可調用,模型 ID gemini-4-carbon;從 Argon 遷移只需改模型名,請求結構不變。免費層沒有 Carbon,需要綁定賬單賬戶並預付。
Google AI Ultra 訂閱用戶在 Gemini App、Search AI Mode、Antigravity 中使用,模型選擇器提供 Low / Medium / High 三檔(默認 High)。Free、Plus、Pro 訂閱均無。
企業路線:API 或下拉菜單調用,Antigravity 與 CodeMender 中可用。
Google Antigravity 中的多 Agent 編排預覽:一個提示詞派出一隊協調 Agent,可把 Carbon 指派給架構、協調、審計角色,搭配更快更輕的模型與對抗驗證器。
OpenRouter 上爲 google/gemini-4-carbon,與官方同價 $2/$10,:batch 變體 $1/$5;models.dev 目錄有 16 家提供商列出 Carbon。Cursor 在 11 月 10 日當天接入。
科學家:10,000 名科學家項目入選者免費 6 個月在 Antigravity 最高用量檔用 Argon + Carbon,申請開放、每週滾動入選。網絡安全:專項配置仍通過 Fairwind Program 申請。
單價與 Argon 完全相同,但賬單未必相同——Carbon 每個任務花的思考 token 更多。價格爲 2026 年 11 月 11–12 日覈查口徑,結賬頁價格爲準。
| API 檔位(每百萬 tokens) | 至 2026-11-30 | 2026-12-01 起 |
|---|---|---|
| Standard 輸入 | $2.00 | $4.00 |
| Standard 輸出(含思考 tokens) | $10.00 | $20.00 |
| 緩存輸入(隱式緩存,95% 折扣) | $0.10 | $0.20 |
| Batch API(輸入 / 輸出) | $1 / $5 | $2 / $10 |
| Flex(輸入 / 輸出) | $1 / $5 | $2 / $10 |
| Priority(輸入 / 輸出) | $3.60 / $18 | $7.20 / $36 |
緩存爲隱式緩存:提示不少於 4,096 tokens 時自動觸發。Batch / Flex / Priority 價格引自定價頁梳理(agentdeals.dev,11 月 12 日)。
美國定價 $99.99/月(Ultra 5x,Pro 用量 5 倍)/ $199.99/月(Ultra 20x)。Ultra 在 150+ 國家銷售;Carbon 覆蓋 Ultra 用戶的「部分國家」,Google 未列舉具體名單。Pro、Plus、免費訂閱均無 Carbon——這一分層在 11 月 10–11 日的 Instagram / Facebook 評論中引發付費 Pro 用戶不滿,不少人拿 Claude Pro 提供 Fable 5.2 作對比。
Artificial Analysis 按標準價測得:Carbon(High)每任務 $7.76,約比 Claude Opus 5.5(max,$10.04)低 23%,約 1.2 倍於 GPT-6 Astra(max,$6.25);Fable 5.2(High)爲 $9.08,只比 Carbon 貴 1 美元多。分檔看:Medium 每任務 $6.59 反而比 Low($7.19)便宜——弱檔位更容易繞圈子、耗更多 token。簡單任務上 Carbon 明顯更貴:同一組 11 個小任務,Carbon 默認檔成本是 Argon 的 3.05 倍;但在編碼 Agent 任務上,Carbon 爲 $7.74/任務,反而低於 Argon 的 $11.73。
以下全部是 11 月 10–12 日的首批上手實測:樣本小、方法不一,請按「個案」看待。每條都給出結論徽章;點「閱讀本站全文」進入本站的深度改寫文章頁,文末附原始出處。
來源:Quantslant,2026-11-11(vs Argon、Fable 5.2、GPT-6.1 Sol,同設置、單元測試評分)
23 項實用任務 Carbon 99.5%($0.23,264 秒),與 Argon 持平但慢一倍;Agent、電子表格(34/34)、安全審查(8/8)、SOC 日誌(7/7)全部滿分。但 6 項編碼任務、76 個單元測試中 Carbon 僅 63/76,三位對手全部 76/76。另有思考燒光 Bug:正則任務 medium 檔用掉全部 16,000 tokens 推理,未返回代碼。
閱讀本站全文 → (本站原創整理)
來源:SeedRouter / blog.laozhang.ai,2026-11-11
同提示詞測試:兩者答案都對,但 Carbon 計費輸出 token 多約 44%(小樣本);11 個小任務上 Carbon 成本是 Argon 的 3.05 倍、耗時 2.75 倍。反過來:1,500 行日誌精確計數,Argon 22 秒答錯,Carbon 214 秒答對;編碼 Agent 任務上 Carbon 每任務 $7.74,反而低於 Argon 的 $11.73。Google 的建議:交互式編程用 Argon,Argon 答錯時換 Carbon。
閱讀本站全文 → (本站原創整理)
來源:eesel.ai / Inferera,2026-11-11/12
AA 三檔數據:Low 54 分、$7.19/任務、編造率 20%;Medium 58 分、$6.59/任務、編造率 4.2%(最便宜也最誠實);High 60 分、$7.76/任務、編造率 11%。代價是時間:首個詞約等 1 分鐘,high 檔每基準任務約 26 分鐘。API 側:thinking_budget、minimal 檔、candidate_count>1 均返回 400;低輸出上限會靜默截斷思考、返回空文本且照樣收費;建議 HTTP 超時 15–20 分鐘。
閱讀本站全文 → (本站原創整理)
來源:Google 官方 10 月 28 日分數表;Artificial Analysis;LMArena,2026-11-07/12
Google 官方 17 項表:Carbon 15 項第一,只輸 FrontierSWE 2(61.2%,低於 Astra 的 65.5%)與 OSWorld-2.1(81.5%,低於 Opus 5.5 的 87.2%)。但多項分數由 Google 自行計算,安全測試在無網、防護關閉環境下進行,且對比的是上一代 Fable 5.1。獨立榜單:AA 指數 60(11 月 7 日登頂,11 月 9 日被 Fable 5.2 的 65 超過)、LMArena 1526 預覽分第一、WebDev Arena 第 4。
閱讀本站全文 → (本站原創整理)
來源:虎嗅,2026-11-11(Antigravity,High 檔,全部一次成型)
9 個任務沒有一項徹底失敗:戶型圖變可漫遊 3D 樓棟(約 1 小時)、iPhone Duo 交互 3D 展品、手寫布料物理絲綢卷軸、逐幀復刻 20 秒宣傳片(37 個硬切與原片一致)、8D 重慶夜景等。短板是精細度:傢俱簡單、材質重複、屏幕內容超出圓角機身;15 秒藝術史第一版還「作弊」調用了圖像生成。作者認爲 Google 不再宣稱「最強」,轉而強調任務成本不到第二名一半、價格約 Fable 5.2 的四分之一。
閱讀本站全文 → (本站原創整理)
來源:騰訊雲開發者社區,2026-11-11(Cursor,3 個並行子 Agent,High 檔,複用 Fable 5.2 同題)
「上古卷軸」被服務端錯誤 3 次打斷,耗時 2 小時 16 分鐘(Fable 5.2 爲 86 分鐘、零中斷);重慶 8D 城市生成器 34 分鐘交卷但只有 184 棟方盒子樓(Fable 5.2 爲 1,230 棟);溶洞項目第一輪近 2 小時後作廢重來,大黃狗是圓柱拼的、鏡頭會穿進巖壁。3 個項目共花 $21.30,遠低於 Fable 5.2 的 500 多塊;作者結論:「不要完全相信跑分」。
閱讀本站全文 → (本站原創整理)
以下 6 篇均爲本站原創改寫與再整理的深度文章——重新組織結構、補充結論與選型建議,事實與數據與原始測評覈對一致。點擊任意一篇在本站內閱讀,文末附原始出處。
Carbon 與 Qwen4-Max 相反:分數很多,問題是誰測的、跟誰比。下面先給 Google 官方 17 項表(2026-10-28),再給獨立榜單,最後說 caveat。
| 測試(Google 官方表) | Carbon | Argon | GPT-6 Astra | Opus 5.5 | Fable 5.1 |
|---|---|---|---|---|---|
| AA Index v5.0 | 59 | 50 | 57 | 58 | 49 |
| Vals Index v2.2 | 70.4% | 69.2% | 65.0% | 70.2% | — |
| SWE-sweep | 9.6% | 4.3% | 9.5% | 3.6% | 2.9% |
| FrontierSWE 2 | 61.2% | 55.0% | 65.5% | 62.3% | 56.3% |
| Terminal-Bench 4.0 (AA) | 65.2% | 57.1% | 59.1% | 59.6% | 52.0% |
| Agents' Last Exam v1.1 | 77.4% | 69.0% | 73.1% | 71.6% | — |
| OSWorld-2.1(離線) | 81.5% | 74.7% | 80.9% | 87.2% | 76.8% |
| AutomationBench (Zapier) | 53.9% | 51.3% | 41.4% | 42.5% | 31.4% |
| Vending-Bench 2(美元餘額) | 16,484 | 13,718 | 15,515 | 9,235 | 5,422 |
| Chartography | 72.5% | 71.6% | 71.0% | 66.3% | 46.2% |
| Blueprint-Bench 2 | 58.6% | 54.4% | 49.7% | 51.2% | 41.9% |
| GraphWalks 256K–1M | 96.2% | 84.2% | 71.8% | 66.8% | 65.0% |
| RiemannBench | 80.0% | 76.0% | 72.0% | 69.6% | 65.6% |
| Terminal-Bench-Science 0.1 (AA) | 64.8% | 52.9% | 63.3% | 59.0% | 43.3% |
| CWE-bench | 71.0% | 68.0% | 68.0% | 67.0% | 58.0% |
| ExploitGym | 54.8% | 42.0% | 42.4% | 34.5% | — |
| SEC-bench Pro | 88.2% | 83.4% | 85.4% | — | — |
Carbon 15/17 第一。官方表對比的是上一代 Claude Fable 5.1,而非 11 月 9 日發佈的 Fable 5.2。
| 測量 | 來源 | Carbon 成績 | 對照 |
|---|---|---|---|
| AA 智能指數 v5.0.1 | Artificial Analysis,11 月 7 日 | 60.0(High),曾登頂 | Fable 5.2 65(11 月 9 日起居首);Opus 5.5 59;Astra 57;Argon 50 |
| AA 指數分檔 | Artificial Analysis / eesel,11 月 11 日 | Low 54 / Medium 58.3 / High 60 | Medium 每任務 $6.59 最便宜,編造率 4.2% 最低 |
| AA-Omniscience v1.1 | Artificial Analysis,約 10 月 31 日 | 淨分 54,第 1 | Opus 5.5 48;準確率 58%(Argon 48%);幻覺率 11%(Argon 4%) |
| LMArena Text | LMArena,11 月 6–7 日 | 1526 預覽分,第 1(約 4,483 票) | Argon(High)1523;Muse Wonder 1513;Opus 5.5 1505 |
| WebDev Arena | LMArena | 1768 elo,第 4 | — |
| CursorBench 4.0 | Cursor | 第 1(high 檔 60.4%) | — |
| Coding Agent Index | Artificial Analysis | 59 | Claude Code + Opus 5.5 爲 62,Carbon 落後 3 分 |
| MKA(多模態知識 Agent) | Google × AA 合作新基準 | 第 2 | 平均任務成本不到第一名的 1/6 |
| FrontierSWE v2 | Snowtraces,11 月 11 日 | 61.2% | Fable 5.2 69.9% |
完整盤點見本站基準測試專文 →
| 維度 | Gemini 4 Carbon | Gemini 4 Argon |
|---|---|---|
| 狀態 | GA(2026-11-10) | GA(更早發佈) |
| API 價格 | $2 / $10 至 11 月 30 日,之後 $4 / $20 | 與 Carbon 完全相同 |
| 上下文 | 1M | 1M |
| AA 指數 | 60(High)/ 58(Medium)/ 54(Low) | 50(High) |
| 速度 | 慢:每任務約 26–27 分鐘,首 token 約 68 秒 | 快得多:同組任務約一半時間 |
| 訂閱 | 僅 Ultra | Pro 及以上可用 |
| 適合 | 最難的任務:長程 Agent、疑難調試、科學推理、安全審查 | 大多數編程與知識工作、交互式編程、快速循環 |
Google 自己的建議:Argon 仍是大多數軟件工程與知識工作的正確選擇;Carbon 留給最難的任務。
| 模型 | API 價格 $/百萬(入/出) | AA 指數 | 上下文 |
|---|---|---|---|
| Claude Fable 5.2 | 8 / 40 | 65(max) | 1M |
| Gemini 4 Carbon | 2 / 10(至 11 月 30 日) | 60(high) | 1M |
| Claude Opus 5.5 | — | 59(max) | — |
| GPT-6 Astra | 10 / 50 | 57(max) | 1.05M |
| GPT-6.1 Sol | 2 / 10 | 55(max) | 1.05M |
| Qwen3.8 Max | 2 / 6 | 45(09-27) | 1M |
| Qwen4-Max-Preview | 僅 Token Plan 訂閱 | 未收錄 | 1M |
Carbon 沒有出現大規模謠言潮,但以下說法正在流傳或容易混淆,需要逐條澄清。
都是真的,只是版本不同:59 是 AA 指數 v5.0(10 月 28 日公告時),60 是 v5.0.1(11 月 7 日更新後)。引用時請註明版本。
曾是,但只維持約兩天(11 月 7–9 日)。11 月 9 日 Claude Fable 5.2 發佈後以 65 分居首,截至 11 月 12 日 Carbon 60 分居其後。另注意:不同統計口徑(按模型 / 按配置)給出的 Carbon 排名在 2–4 位之間,本頁不寫單一排名數字。
AA 指數(65 對 60)、FrontierSWE v2(69.9% 對 61.2%)均爲 Fable 5.2 領先。Google 官方對比表用的是上一代 Fable 5.1——這張表發佈時 Fable 5.2 還未發佈,但現在拿它論證「Carbon 最強」已經過時。
Google 自述 Carbon 基於 Argon 的預訓練基礎與架構,「主要差異是每輪花多少思考 token」,且單價完全相同。但這不等於「同一個模型」:額外後訓練帶來了實質改進——作弊嘗試 3.6%→0.4%(成功 2.4%→0%)、Agent 誠實基準 70%→92%、AA 指數 50→60。
定價頁明確 Carbon 不在免費層;Google AI Pro 訂閱同樣沒有,Pro 仍爲 Argon。消費者側只有 Ultra 訂閱能用。
2026 年 10 月 28 日宣佈,11 月 10 日全面開放(GA)。10 月 28 日至 11 月 10 日之間僅開放給 Fairwind Program 的部分網絡安全防禦方。時間線見本頁第 02 節。
不免費。API 沒有免費層;消費者側只有 Google AI Ultra 訂閱($99.99 / $199.99 每月,美國)能用,Pro、Plus、免費訂閱均無。科學家計劃入選者可免費使用 6 個月。
至 11 月 30 日:輸入 $2 / 輸出 $10(每百萬 tokens,含思考 tokens);12 月 1 日起翻倍至 $4 / $20。緩存輸入 95% 折扣($0.10 → $0.20)。Batch 與 Flex 爲半價。價格見本頁第 05 節。
AA 指數 60 對 50,安全與誠實指標明顯更好;但日常任務上它做不了 Argon 做不了的事,還更慢更貴。Google 的建議:交互式編程與大多數知識工作用 Argon;Argon 答錯或漏東西時換 Carbon。詳見本站Carbon vs Argon 對比。
按分數不強:AA 指數 65 對 60、FrontierSWE v2 69.9% 對 61.2%,均爲 Fable 5.2 領先;前端實測中 Fable 5.2 的審美與穩定性也更好。但 Carbon 單價只有 Fable 5.2 的四分之一,任務成本 $7.76 對 $9.08,差距只有 1 美元多。
默認 medium 是最划算的檔:AA 58 分、每任務 $6.59、編造率 4.2%(三檔中最低);High 60 分但每任務 $7.76、編造率 11%;Low 54 分、每任務 $7.19、編造率 20%。詳見本站思考檔位與避坑指南。
輸出上限覆蓋思考 token:上限設太低時,模型把 token 全花在思考上、還沒寫答案就觸頂,返回空文本、狀態 incomplete,費用照收。這是已確認的行爲(Quantslant 實測、Google 文檔均有說明),不是偶發故障。解決辦法:把 max_output_tokens 設到至少 32,768,用思考檔位而不是輸出上限控成本。
1,048,576 tokens,輸入、思考與輸出共享。最大輸出 65,536 tokens(流式)或 32,768(非流式)。
不是開源模型,沒有發佈權重,HuggingFace 上沒有官方模型。只能通過 Google API 與訂閱服務使用。
GA 公告材料未說明 Vertex AI 的可用性,截至 11 月 13 日沒有確認消息。請以 Google Cloud 官方文檔爲準。
Google 官方博客(blog.google)、Gemini API 定價頁與模型頁、Artificial Analysis 模型頁,以及本頁下方的來源列表。
本頁事實均來自以下公開來源(覈查於 2026 年 11 月 12–13 日)。
Google's strongest model, generally available since November 10, 2026: built for long-horizon agent work, complex coding, cybersecurity and scientific reasoning. The API costs the same as Argon ($2/$10 through November 30), but it spends more thinking tokens and runs slower; its Artificial Analysis index of 60 briefly topped the chart before Claude Fable 5.2 passed it two days later. This page collects the official specs, the ways to try it, pricing, and the first hands-on reviews in English and Chinese — with unverified claims clearly marked.
What is confirmed: this is a general release, not a time-boxed preview — but its "strongest" title lasted about two days.
gemini-4-carbon (stable API identifier)Thirteen days from announcement to GA; about two days at the top of the AA index.
Google announces Gemini 4 Carbon on blog.google; same-day access limited to a subset of cybersecurity defenders via the Fairwind Program (the specialized cyber configuration still requires an application). The announcement claims AA index v5.0 score of 59, ranked first.
LMArena Text preliminary score around 1528, provisionally #1 (~4,403 votes).
Artificial Analysis index updates to v5.0.1: Carbon (High) scores 60 and takes the top spot; Opus 5.5 at 59, Astra at 57.
Claude Fable 5.2 ships, scores 65 on the AA index and takes #1 from Carbon; it also leads FrontierSWE v2 at 69.9% vs Carbon's 61.2%.
General availability: paid API, AI Studio, Gemini for business, Ultra subscribers (Gemini app, AI Mode, Antigravity); model card published; Antigravity launches the /swarm multi-agent preview; Cursor integrates Carbon the same day.
First hands-on wave: Quantslant, Huxiu, Yupi, SeedRouter, eesel and others publish reviews in English and Chinese (see section 06).
Introductory API pricing ends ($2/$10).
API prices double to $4 input / $20 output per 1M tokens.
Official figures from the model page and model card (checked Nov 11, 2026); unpublished items are stated as such, never filled in by guesswork.
Developers: the paid API is fastest. Everyone else: only the Ultra subscription — Pro doesn't include Carbon.
The main developer route. Any paid-tier API key can call it as gemini-4-carbon; migrating from Argon means changing only the model string. There is no free tier — you need a billing account with prepaid credit.
Google AI Ultra subscribers get Carbon in the Gemini app, AI Mode in Search and Antigravity, with Low / Medium / High thinking levels in the model picker (High by default). Free, Plus and Pro plans don't get it.
The enterprise route: API or dropdown access, available in Antigravity and CodeMender.
Multi-agent orchestration preview in Google Antigravity: one prompt dispatches a coordinated team of agents; assign Carbon to architecture, coordination and auditing roles, paired with faster, lighter models and adversarial verifiers.
On OpenRouter as google/gemini-4-carbon at the same $2/$10 price, with a :batch variant at $1/$5; 16 providers list Carbon in the models.dev catalog. Cursor integrated it on November 10.
Scientists: selected participants in the expanded 10,000-scientists program get Argon + Carbon free for 6 months in Antigravity's highest usage tier; applications are open with weekly rolling onboarding. Cybersecurity: the specialized configuration is still gated through Fairwind Program applications.
Same per-token price as Argon — but not necessarily the same bill, because Carbon spends more thinking tokens per task. Prices checked Nov 11–12, 2026; the checkout page is authoritative.
| API tier (per 1M tokens) | Through Nov 30, 2026 | From Dec 1, 2026 |
|---|---|---|
| Standard input | $2.00 | $4.00 |
| Standard output (thinking tokens included) | $10.00 | $20.00 |
| Cached input (implicit caching, 95% off) | $0.10 | $0.20 |
| Batch API (in / out) | $1 / $5 | $2 / $10 |
| Flex (in / out) | $1 / $5 | $2 / $10 |
| Priority (in / out) | $3.60 / $18 | $7.20 / $36 |
Caching is implicit: it switches on automatically for prompts of at least 4,096 tokens. Batch / Flex / Priority rates are from a pricing-page summary (agentdeals.dev, Nov 12).
US pricing: $99.99/month (Ultra 5x, 5x Pro's limits) / $199.99/month (Ultra 20x). Ultra is sold in 150+ countries; Carbon reaches Ultra subscribers in "selected countries" that Google has not enumerated. Pro, Plus and free plans don't include Carbon — a tiering choice that drew complaints from paying Pro subscribers in Instagram and Facebook comments on Nov 10–11, many comparing it unfavourably with Claude Pro offering Fable 5.2.
Artificial Analysis at list price: Carbon (High) costs $7.76 per task, about 23% less than Claude Opus 5.5 (max, $10.04) and about 1.2x GPT-6 Astra (max, $6.25); Fable 5.2 (High) is $9.08 — only about a dollar more than Carbon. By thinking level, Medium at $6.59 per task is actually cheaper than Low ($7.19), because weaker settings loop more and burn more tokens. On easy work Carbon is clearly pricier: on one set of 11 small tasks, Carbon at default settings cost 3.05x Argon; but on coding-agent tasks Carbon came in at $7.74 per task against Argon's $11.73.
Official pages: Gemini API pricing · Google's Nov 10 GA announcement
All from Nov 10–12: small samples, differing methods — read each as a case study. Every card carries a verdict badge; "Read our full write-up" opens this site's in-depth rewrite, with the original source linked at the end.
Source: Quantslant, Nov 11, 2026 (vs Argon, Fable 5.2, GPT-6.1 Sol; same settings, unit-test grading)
On 23 practical tasks Carbon scored 99.5% ($0.23, 264s) — tied with Argon but twice as slow; agent, spreadsheet (34/34), security review (8/8) and SOC log analysis (7/7) were all perfect. But on 6 coding tasks with 76 unit tests, Carbon passed only 63/76 while all three rivals passed 76/76. There was also a thinking-burn bug: a regex task at medium effort spent all 16,000 tokens on reasoning and returned no code.
Read our full write-up → (original reporting by this site)
Source: SeedRouter / blog.laozhang.ai, Nov 11, 2026
Same-prompt tests: both models got the answers right, but Carbon billed about 44% more output tokens (small sample); on 11 small tasks Carbon cost 3.05x Argon and took 2.75x as long. The other way round: on exact counting over a 1,500-line log, Argon answered wrong in 22 seconds while Carbon answered right in 214; on coding-agent tasks Carbon cost $7.74 per task against Argon's $11.73. Google's guidance: Argon for interactive coding, Carbon when Argon's answer misses something.
Read our full write-up → (original reporting by this site)
Source: eesel.ai / Inferera, Nov 11–12, 2026
AA data by level: Low scores 54 at $7.19/task with a 20% made-up-answer rate; Medium scores 58 at $6.59/task with 4.2% made up (cheapest and most honest); High scores 60 at $7.76/task with 11% made up. The cost is time: about a minute before the first word, ~26 minutes per benchmark task on high. On the API side, thinking_budget, the minimal level and candidate_count>1 all return 400; a low output cap silently truncates thinking and returns empty text that is still billed; HTTP timeouts of 15–20 minutes are recommended.
Read our full write-up → (original reporting by this site)
Source: Google's official Oct 28 score table; Artificial Analysis; LMArena, Nov 7–12, 2026
Google's official 17-test table: Carbon first in 15, losing only FrontierSWE 2 (61.2%, below Astra's 65.5%) and OSWorld-2.1 (81.5%, below Opus 5.5's 87.2%). But Google calculated many of the scores itself, security tests ran offline with safeguards disabled, and the comparison used the previous-generation Fable 5.1. Independent boards: AA index 60 (topped the chart Nov 7, passed by Fable 5.2's 65 on Nov 9), LMArena 1526 preliminary #1, WebDev Arena #4.
Read our full write-up → (original reporting by this site)
Source: Huxiu, Nov 11, 2026 (Antigravity, High level, all one-shot)
None of the 9 tasks failed outright: a floor plan became a walkable 3D building (~1 hr), an iPhone Duo interactive 3D exhibit, hand-written cloth-physics silk scroll, a frame-by-frame remake of a 20-second promo (37 hard cuts matching the original), an 8D Chongqing night scene and more. The shortfall is finish: simple furniture, repeated materials, screen content poking out of a rounded body; the 15-second art-history piece first "cheated" by calling image generation. The author notes Google no longer claims "strongest" and instead stresses a cost per task under half of #2's, at about a quarter of Fable 5.2's price.
Read our full write-up → (original reporting by this site)
Source: Tencent Cloud Developer Community, Nov 11, 2026 (Cursor, 3 parallel sub-agents, High level, same prompts as the Fable 5.2 test)
The "ancient scroll" project was interrupted 3 times by server errors and took 2 hours 16 minutes (Fable 5.2: 86 minutes, zero interruptions); the Chongqing 8D city generator shipped in 34 minutes but with only 184 boxy buildings (Fable 5.2: 1,230); the cave project had to be scrapped and restarted after nearly 2 hours, and its dog was cylinders with detached legs while the camera clipped through rock. All three projects cost $21.30, far below Fable 5.2's 500+ RMB; the author's conclusion: "don't fully trust benchmarks."
Read our full write-up → (original reporting by this site)
All 6 articles below are original rewrites and re-analysis by this site — restructured, with added conclusions and selection advice, facts and figures checked against the original reviews. Open any of them to read on-site; the original source is linked at the end of each.
The opposite of Qwen4-Max: plenty of scores — the question is who ran them and against whom. First Google's official 17-test table (Oct 28, 2026), then the independent boards, then the caveats.
| Test (Google's official table) | Carbon | Argon | GPT-6 Astra | Opus 5.5 | Fable 5.1 |
|---|---|---|---|---|---|
| AA Index v5.0 | 59 | 50 | 57 | 58 | 49 |
| Vals Index v2.2 | 70.4% | 69.2% | 65.0% | 70.2% | — |
| SWE-sweep | 9.6% | 4.3% | 9.5% | 3.6% | 2.9% |
| FrontierSWE 2 | 61.2% | 55.0% | 65.5% | 62.3% | 56.3% |
| Terminal-Bench 4.0 (AA) | 65.2% | 57.1% | 59.1% | 59.6% | 52.0% |
| Agents' Last Exam v1.1 | 77.4% | 69.0% | 73.1% | 71.6% | — |
| OSWorld-2.1 (offline) | 81.5% | 74.7% | 80.9% | 87.2% | 76.8% |
| AutomationBench (Zapier) | 53.9% | 51.3% | 41.4% | 42.5% | 31.4% |
| Vending-Bench 2 (USD balance) | 16,484 | 13,718 | 15,515 | 9,235 | 5,422 |
| Chartography | 72.5% | 71.6% | 71.0% | 66.3% | 46.2% |
| Blueprint-Bench 2 | 58.6% | 54.4% | 49.7% | 51.2% | 41.9% |
| GraphWalks 256K–1M | 96.2% | 84.2% | 71.8% | 66.8% | 65.0% |
| RiemannBench | 80.0% | 76.0% | 72.0% | 69.6% | 65.6% |
| Terminal-Bench-Science 0.1 (AA) | 64.8% | 52.9% | 63.3% | 59.0% | 43.3% |
| CWE-bench | 71.0% | 68.0% | 68.0% | 67.0% | 58.0% |
| ExploitGym | 54.8% | 42.0% | 42.4% | 34.5% | — |
| SEC-bench Pro | 88.2% | 83.4% | 85.4% | — | — |
Carbon first in 15 of 17. The official table compares against the previous-generation Claude Fable 5.1, not Fable 5.2 (released Nov 9).
| Measurement | Source | Carbon's result | Comparison |
|---|---|---|---|
| AA Intelligence Index v5.0.1 | Artificial Analysis, Nov 7 | 60.0 (High), briefly #1 | Fable 5.2 65 (#1 since Nov 9); Opus 5.5 59; Astra 57; Argon 50 |
| AA index by level | Artificial Analysis / eesel, Nov 11 | Low 54 / Medium 58.3 / High 60 | Medium cheapest at $6.59/task, lowest made-up rate at 4.2% |
| AA-Omniscience v1.1 | Artificial Analysis, ~Oct 31 | Net score 54, #1 | Opus 5.5 48; accuracy 58% (Argon 48%); hallucination rate 11% (Argon 4%) |
| LMArena Text | LMArena, Nov 6–7 | 1526 preliminary, #1 (~4,483 votes) | Argon (High) 1523; Muse Wonder 1513; Opus 5.5 1505 |
| WebDev Arena | LMArena | 1768 elo, #4 | — |
| CursorBench 4.0 | Cursor | #1 (60.4% at high effort) | — |
| Coding Agent Index | Artificial Analysis | 59 | Claude Code + Opus 5.5 at 62 — Carbon trails by 3 |
| MKA (Multimodal Knowledge Agent) | Google × AA, new benchmark | #2 | Average cost per task under 1/6 of the top scorer's |
| FrontierSWE v2 | Snowtraces, Nov 11 | 61.2% | Fable 5.2 69.9% |
Full roundup: our benchmarks article →
| Dimension | Gemini 4 Carbon | Gemini 4 Argon |
|---|---|---|
| Status | GA (Nov 10, 2026) | GA (earlier) |
| API price | $2 / $10 through Nov 30, then $4 / $20 | Identical to Carbon |
| Context | 1M | 1M |
| AA index | 60 (High) / 58 (Medium) / 54 (Low) | 50 (High) |
| Speed | Slow: ~26–27 min per task, ~68s to first token | Much faster: roughly half the time on the same tasks |
| Subscription | Ultra only | Pro and above |
| Best for | The hardest jobs: long-horizon agents, tough debugging, science reasoning, security review | Most coding and knowledge work, interactive coding, fast loops |
Google's own advice: Argon remains the right choice for most software engineering and knowledge work; keep Carbon for the hardest tasks.
Read the deep comparison: Carbon vs Argon →
| Model | API price $/1M (in/out) | AA index | Context |
|---|---|---|---|
| Claude Fable 5.2 | 8 / 40 | 65 (max) | 1M |
| Gemini 4 Carbon | 2 / 10 (through Nov 30) | 60 (high) | 1M |
| Claude Opus 5.5 | — | 59 (max) | — |
| GPT-6 Astra | 10 / 50 | 57 (max) | 1.05M |
| GPT-6.1 Sol | 2 / 10 | 55 (max) | 1.05M |
| Qwen3.8 Max | 2 / 6 | 45 (Sep 27) | 1M |
| Qwen4-Max-Preview | Token Plan only | Not listed | 1M |
No major rumor wave has formed around Carbon, but these claims are circulating or easy to get wrong.
Both are real — different index versions: 59 on AA index v5.0 (at the Oct 28 announcement), 60 on v5.0.1 (after the Nov 7 update). Cite the version when you quote it.
It was, for about two days (Nov 7–9). Claude Fable 5.2 shipped on Nov 9 and leads at 65; as of Nov 12 Carbon at 60 sits behind it. Note also that different counting methods (by model vs by configuration) place Carbon anywhere from 2nd to 4th, so this page doesn't print a single rank number.
Fable 5.2 leads on the AA index (65 vs 60) and FrontierSWE v2 (69.9% vs 61.2%). Google's official comparison table used the previous-generation Fable 5.1 — it predates Fable 5.2, so citing it to claim Carbon is strongest is now out of date.
Google itself says Carbon is built on Argon's pre-training foundation and architecture, that "the main difference is how many thinking tokens each model spends per turn", and charges the same per-token price. But it is not the same model: the additional post-training produced real improvements — cheating attempts 3.6% → 0.4% (successful 2.4% → 0%), agentic honesty benchmark 70% → 92%, AA index 50 → 60.
The pricing page explicitly excludes Carbon from the free tier; Google AI Pro doesn't include it either — Pro still gets Argon. On the consumer side, only Ultra subscribers get Carbon.
Announced October 28, 2026, and generally available November 10, 2026. Between those dates it was limited to a subset of cybersecurity defenders in the Fairwind Program. See the timeline in section 02.
No. The API has no free tier; on the consumer side only Google AI Ultra subscribers ($99.99 / $199.99 per month, US) get it — Pro, Plus and free plans don't. Selected scientists get 6 months free through the scientists program.
Through November 30: $2 input / $10 output per 1M tokens (thinking tokens included); from December 1 the price doubles to $4 / $20. Cached input is 95% off ($0.10 → $0.20). Batch and Flex are half price. See section 05.
AA index 60 vs 50, and clearly better safety and honesty metrics; but on everyday tasks it does nothing Argon can't, more slowly and at a higher bill. Google's advice: Argon for interactive coding and most knowledge work; switch to Carbon when Argon's answer misses something. See our Carbon vs Argon comparison.
Not by the scores: AA index 65 vs 60 and FrontierSWE v2 69.9% vs 61.2% both favour Fable 5.2, and front-end tests also rate Fable 5.2's taste and stability higher. But Carbon costs a quarter of Fable 5.2's per-token price, and the cost per task is $7.76 vs $9.08 — only about a dollar apart.
Medium, the default, is the best value: AA 58, $6.59 per task, a 4.2% made-up-answer rate (lowest of the three). High scores 60 but costs $7.76 per task with an 11% made-up rate; Low scores 54 at $7.19 per task with a 20% made-up rate. See our thinking levels and pitfalls guide.
The output cap covers thinking tokens: set it too low and the model spends everything on reasoning, hits the cap before writing an answer, and returns empty text with status incomplete — billed anyway. This is confirmed behaviour (Quantslant measured it; Google's docs describe it), not a random glitch. Fix: set max_output_tokens to at least 32,768 and control cost with the thinking level instead of the output cap.
1,048,576 tokens, shared by input, thinking and output. Max output is 65,536 tokens (streaming) or 32,768 (non-streaming).
No. No weights have been published and there is no official model on HuggingFace. It is available only through Google's API and subscription services.
The GA materials do not address Vertex AI availability, and there was no confirmed news as of November 13. Check Google Cloud's official documentation.
Google's official blog (blog.google), the Gemini API pricing and model pages, the Artificial Analysis model page, and the source list below.
Facts on this page come from the public sources below (checked Nov 12–13, 2026).