程序员鱼皮在 11 月 11 日发布了 Carbon 的中文前端实测，标题直截了当：「失望透顶」。测试方法值得肯定：复用前一天 Fable 5.2 测试的同一 3 道题、提示词一字不改，在 Cursor 里同时开 3 个子 Agent 并行跑，全部选 Carbon 的 High 档，全程不做人工干预。这是最接近「同场竞技」的一次中文对比。先说结论：Carbon 在审美细节与稳定性上与 Fable 5.2「差了不止一星半点」，但 3 个项目只花了 $21.30，不到 Fable 5.2 的三分之一。
要求把一篇分 6 个阶段的长文做成古籍设计感的网页，画面要顶级、经得起放大细看，打开要有展开古物的仪式感。执行中 AI 先后 3 次被服务端错误 [unavailable] Error 打断，作者排查后发现每次都发生在 AI 想一口气写出一个超大文件时（比如正文组件一个文件就有 68KB）。从派发到出成品截图，足足折腾了 2 小时 16 分钟——Fable 5.2 跑这道题花了 86 分钟，一次都没中断过。成品本身不差：暗色背景上的卷轴、红色「启」字火漆印、「鱼皮亲撰」方印、六张带线描徽记的阶段卡片（还起了「启蒙之境」「筑基之境」这种修仙味的名字）。但与 Fable 5.2 并排对比，差别明显：Fable 5.2 在黑暗里打了一束光在青绿锦面卷轴上，仪式感更强；Carbon 信息更多，但看着更像常规网页，古物的陈旧感弱一些。
要求对标 OpenAI 官方 Showcase 的程序化城市生成器，把重庆的层叠立交、轻轨穿楼、依山高差做出来。这是 Carbon 表现最好的一题：首次构建零报错、自己修改 2 轮，34 分钟就交卷——Fable 5.2 花了 88 分钟，Carbon 快了一倍多。功能也很全：9 个机位、5 套预设、4 套配色，调整参数实时生效，可一键更换随机种子；李子坝机位里轨道确实从楼中间穿过去。但第一眼的问题也很明显：浑黄的长江、碧绿的嘉陵江、红色斜拉桥都有，楼群却整整齐齐挤在一块方形底板上。数字对比最能说明差距：Fable 5.2 用 1,230 栋楼铺满了两江半岛，Carbon 只有 184 栋方盒子楼——虽然比 12 分钟糊弄出灰绿色方盒子的 GPT-6 Sol 强不少，但精细度还是差了一个档次。
难度最高的一题：用纯代码捏不规则的自然溶洞，再放一只 3D 仿真大黄狗（中华田园犬）在洞里穿梭，禁用任何现成模型与贴图。结果最折腾：第一轮跑了将近 2 小时，中途被服务端错误打断两次，最后只留下 5 个文件、2,000 多行代码，连入口页面都没有，项目根本打不开，只能作废重来。原因与第一题相同：AI 想一次性生成整个溶洞场景的大文件导致崩掉。重跑时作者没有改动提示词，只在任务说明里加了一句「单个文件不超过 600 行、分多次写」，之后 AI 只花了 30 分钟就跑出了能玩的版本——这是一个值得记下的实操教训。成品效果参差：溶洞实景、荧光前廊、发光白色石瀑配水面倒影看着还行；但大黄狗「完全是一堆圆柱拼出来的，四只爪子跟腿是断开的，悬在地上，尾巴是一节一节的弯管」；镜头还会直接穿进岩壁，手机上打开默认就是满屏青绿色。作者说这是与 Fable 5.2「差距最大的一项」。
3 个项目一共花了 $21.30（约 142 元），比 Fable 5.2 消耗的 500 多块便宜了一大截。顺带一提：这篇文章本身也是作者让 Carbon 写的，花了 $2.05（约 14 元），作者评价「没有明显感受到 Carbon 写作能力有多强，还是会有明显的 AI 味儿，有不少细节还是需要人工来调整优化」。作者的最终判断：Carbon「可用可不用，没什么让人惊喜的亮点，在价格上也不如国产模型有优势，稳定性上目前也差了一些（也可能是我自己的问题）」；并再次证明了一点——不要完全相信跑分。
需要提醒的是：服务端错误是 GA 后首日的个案，无法确认是普遍故障还是个例；且「单个文件不超过 600 行」这个约束显著改善了稳定性，说明部分问题出在任务拆分方式而非模型本身。但即便排除中断因素，184 对 1,230 栋楼、圆柱拼狗这类审美与精细度差距，仍然是模型层面的结论。
参考来源：本文为本站原创改写与整理，事实与数据核对自 腾讯云开发者社区（鱼皮）前端三题实测（2026-11-11）
← 返回 Gemini 4 Carbon 专题程序員魚皮在 11 月 11 日發佈了 Carbon 的中文前端實測，標題直截了當：「失望透頂」。測試方法值得肯定：複用前一天 Fable 5.2 測試的同一 3 道題、提示詞一字不改，在 Cursor 裏同時開 3 個子 Agent 並行跑，全部選 Carbon 的 High 檔，全程不做人工干預。這是最接近「同場競技」的一次中文對比。先說結論：Carbon 在審美細節與穩定性上與 Fable 5.2「差了不止一星半點」，但 3 個項目只花了 $21.30，不到 Fable 5.2 的三分之一。
要求把一篇分 6 個階段的長文做成古籍設計感的網頁，畫面要頂級、經得起放大細看，打開要有展開古物的儀式感。執行中 AI 先後 3 次被服務端錯誤 [unavailable] Error 打斷，作者排查後發現每次都發生在 AI 想一口氣寫出一個超大文件時（比如正文組件一個文件就有 68KB）。從派發到出成品截圖，足足折騰了 2 小時 16 分鐘——Fable 5.2 跑這道題花了 86 分鐘，一次都沒中斷過。成品本身不差：暗色背景上的卷軸、紅色「啓」字火漆印、「魚皮親撰」方印、六張帶線描徽記的階段卡片（還起了「啓蒙之境」「築基之境」這種修仙味的名字）。但與 Fable 5.2 並排對比，差別明顯：Fable 5.2 在黑暗裏打了一束光在青綠錦面卷軸上，儀式感更強；Carbon 信息更多，但看着更像常規網頁，古物的陳舊感弱一些。
要求對標 OpenAI 官方 Showcase 的程序化城市生成器，把重慶的層疊立交、輕軌穿樓、依山高差做出來。這是 Carbon 表現最好的一題：首次構建零報錯、自己修改 2 輪，34 分鐘就交卷——Fable 5.2 花了 88 分鐘，Carbon 快了一倍多。功能也很全：9 個機位、5 套預設、4 套配色，調整參數實時生效，可一鍵更換隨機種子；李子壩機位裏軌道確實從樓中間穿過去。但第一眼的問題也很明顯：渾黃的長江、碧綠的嘉陵江、紅色斜拉橋都有，樓羣卻整整齊齊擠在一塊方形底板上。數字對比最能說明差距：Fable 5.2 用 1,230 棟樓鋪滿了兩江半島，Carbon 只有 184 棟方盒子樓——雖然比 12 分鐘糊弄出灰綠色方盒子的 GPT-6 Sol 強不少，但精細度還是差了一個檔次。
難度最高的一題：用純代碼捏不規則的自然溶洞，再放一隻 3D 仿真大黃狗（中華田園犬）在洞裏穿梭，禁用任何現成模型與貼圖。結果最折騰：第一輪跑了將近 2 小時，中途被服務端錯誤打斷兩次，最後只留下 5 個文件、2,000 多行代碼，連入口頁面都沒有，項目根本打不開，只能作廢重來。原因與第一題相同：AI 想一次性生成整個溶洞場景的大文件導致崩掉。重跑時作者沒有改動提示詞，只在任務說明裏加了一句「單個文件不超過 600 行、分多次寫」，之後 AI 只花了 30 分鐘就跑出了能玩的版本——這是一個值得記下的實操教訓。成品效果參差：溶洞實景、熒光前廊、發光白色石瀑配水面倒影看着還行；但大黃狗「完全是一堆圓柱拼出來的，四隻爪子跟腿是斷開的，懸在地上，尾巴是一節一節的彎管」；鏡頭還會直接穿進巖壁，手機上打開默認就是滿屏青綠色。作者說這是與 Fable 5.2「差距最大的一項」。
3 個項目一共花了 $21.30（約 142 元），比 Fable 5.2 消耗的 500 多塊便宜了一大截。順帶一提：這篇文章本身也是作者讓 Carbon 寫的，花了 $2.05（約 14 元），作者評價「沒有明顯感受到 Carbon 寫作能力有多強，還是會有明顯的 AI 味兒，有不少細節還是需要人工來調整優化」。作者的最終判斷：Carbon「可用可不用，沒什麼讓人驚喜的亮點，在價格上也不如國產模型有優勢，穩定性上目前也差了一些（也可能是我自己的問題）」；並再次證明了一點——不要完全相信跑分。
需要提醒的是：服務端錯誤是 GA 後首日的個案，無法確認是普遍故障還是個例；且「單個文件不超過 600 行」這個約束顯著改善了穩定性，說明部分問題出在任務拆分方式而非模型本身。但即便排除中斷因素，184 對 1,230 棟樓、圓柱拼狗這類審美與精細度差距，仍然是模型層面的結論。
參考來源：本文爲本站原創改寫與整理，事實與數據覈對自 騰訊雲開發者社區（魚皮）前端三題實測（2026-11-11）
← 返回 Gemini 4 Carbon 專題Programmer Yupi published a Chinese-language front-end test of Carbon on November 11, headlined — bluntly — "utterly disappointing". The method deserves credit: the same 3 tasks and word-for-word prompts from the previous day's Fable 5.2 test, run in Cursor with 3 parallel sub-agents, all on Carbon's High level, with zero human intervention. It is the closest thing to a head-to-head in Chinese coverage so far. The verdict up front: Carbon trails Fable 5.2 by "more than a little" on aesthetic detail and stability — but all three projects cost $21.30, under a third of Fable 5.2's bill.
The brief: turn a long tutorial in 6 stages into a webpage with the feel of an ancient manuscript — top-tier visuals that hold up under zoom, with a ritual sense of unrolling an artifact on open. During execution the AI was interrupted 3 times by server errors ([unavailable] Error); Yupi traced each one to the model trying to write a huge file in one go (one body component alone was 68KB). From dispatch to finished screenshots took 2 hours 16 minutes — Fable 5.2 did the same task in 86 minutes with zero interruptions. The finished page itself was decent: a scroll on a dark background, a red wax seal with the character for "begin", a "written by Yupi" square seal, six stage cards with line-drawn emblems (and cultivation-novel names like "Realm of Enlightenment" and "Realm of Foundation"). But side by side with Fable 5.2 the difference was clear: Fable 5.2 put a beam of light on a teal brocade scroll in the dark, with far more ceremony; Carbon packed in more information but looked more like a regular webpage, with less of the artifact's aged feel.
The brief: match OpenAI's official showcase procedural city generator, with Chongqing's stacked interchanges, light rail through buildings and hillside elevation changes. This was Carbon's best task: zero errors on the first build, two self-directed revision rounds, delivered in 34 minutes — Fable 5.2 took 88 minutes, so Carbon was more than twice as fast. The feature set was complete: 9 camera positions, 5 presets, 4 colour schemes, live parameter changes, one-click random seed; at the Liziba viewpoint the rail genuinely passes through the building. But the first impression was off: the muddy-yellow Yangtze, the green Jialing River and the red cable-stayed bridge were all there, yet the buildings sat squeezed together on a square base plate. The numbers tell the gap best: Fable 5.2 used 1,230 buildings to fill the two-river peninsula; Carbon had only 184 boxy buildings — well ahead of GPT-6 Sol, which slapped together grey-green boxes in 12 minutes, but still a tier below on detail.
The hardest task: build a 3D cave of irregular natural rock in pure code, with a 3D realistic big yellow dog (a Chinese rural dog) wandering through it — no ready-made models or textures allowed. It was also the most troublesome: the first run took nearly 2 hours, was interrupted twice by server errors, and ended with only 5 files and 2,000+ lines of code — no entry page at all, nothing that would open, so it was scrapped and restarted. The cause was the same as task 1: the AI tried to generate the whole cave scene as huge files and crashed. On the rerun Yupi changed nothing in the prompt except adding one line to the task instructions — "no single file over 600 lines, write in multiple passes" — and the AI produced a playable version in just 30 minutes, a practical lesson worth remembering. The result was mixed: the cave itself, the glowing green entrance hall and the luminous white stone waterfall with water reflections looked fine; but the dog was "entirely cylinders stuck together, its four paws detached from its legs and floating above the ground, its tail a series of bent pipes"; the camera also clipped straight through rock walls, and on a phone the default view was a solid green screen. Yupi called it the biggest gap against Fable 5.2.
All three projects cost $21.30 (about 142 RMB), far less than the 500+ RMB Fable 5.2 consumed. As an aside: the article itself was also written by Carbon, at a cost of $2.05 (about 14 RMB); Yupi's verdict on that was that he "didn't noticeably feel Carbon's writing ability was strong — it still had an obvious AI flavour, with many details needing manual adjustment". His final judgement: Carbon is "usable or not, nothing surprisingly bright; no price advantage over Chinese models, and stability is currently a bit lacking (though that may be my own issue)"; and it proved one point again — don't fully trust benchmarks.
Two caveats: the server errors are a single tester's day-one experience and can't confirm a widespread outage; and the "600 lines per file" constraint improved stability markedly, which suggests part of the problem was task decomposition rather than the model itself. But even setting interruptions aside, the 184-vs-1,230 buildings and the cylinder dog are model-level findings about taste and detail.
Reference source: this article is an original rewrite produced by this site; facts and figures were checked against Tencent Cloud Developer Community (Yupi) three front-end tasks test (Nov 11, 2026)
← Back to the Gemini 4 Carbon topic page