🔶 我们让 Grok 4.5、GPT-5.5 和 Claude 构建了相同的应用

Grok 4.5 本周发布，xAI 称其是迄今为止最智能的模型，“为编码和智能体工作与 Cursor 一同训练”。网络上对新型编码模型最流行的“氛围检验”方式便是构建对决：交给它一个提示，看它生成什么应用，发布视频片段。于是我们也这么做了，而且做得很认真。

我们向 Grok 4.5、GPT-5.5、Claude Opus 4.8 和 Claude Fable 5 给出了完全相同的三条提示：构建一个自包含的 HTML 文件（无库、无网络调用）用于交互式应用。每个模型只给一次机会，没有手把手引导，不做提示调优。然后我们在真实浏览器中加载每个结果，动手测试并记录发生的情况。下面每个片段都是模型的原始输出，你可以点击“亲自试玩”来运行实际生成的应用。

失败处理规则：如果某个应用完全没有渲染，我们允许恰好一次重试，并在用到时会明确告知。

第一轮：一个可打乱和还原的 3D 魔方

构建一个色彩丰富、具有 3D 外观的魔方，带有“打乱”和“还原”按钮。魔方必须可见地动画化旋转。

这题很刁钻。它需要真正的 3D 数学、逐面状态管理以及动画，全都放在一个文件里。

亲自试玩：Grok 4.5 · GPT-5.5 · Opus 4.8 · Fable 5

Claude 兄弟在这一轮表现出色。Opus 4.8 和 Fable 5 均在首次尝试时就生成了色彩正确、动画旋转打乱与还原的纯正 3D 魔方。Grok 4.5 在这里栽了跟头：第一次尝试渲染出了标题和按钮，但魔方完全没出现（一片空白），于是我们用掉了唯一的一次重试机会，第二次尝试才生成了一个干净、鲜艳的 3D 魔方。GPT-5.5 则显得格格不入，它只渲染出一个颜色暗淡的单一表面，几乎看不出色彩，而不是一个完整的魔方。获胜者：Opus 与 Fable，并列第一。

第二轮：粒子重力沙盒

在画布上构建一个交互式粒子重力沙盒：数百个带拖尾的粒子，点击添加一个重型吸引子。让它变得令人着迷。

亲自试玩：Grok 4.5 · GPT-5.5 · Opus 4.8 · Fable 5

每个模型都交付了一个可运行的沙盒，所以这一轮的胜负主要看品味。GPT-5.5 做得最令人着迷：发光的霓虹吸引子加上密集、旋转的彩色拖尾。Grok 4.5 走的是干净轨道风，整齐的吸引子环和色彩斑斓的粒子轨迹。Fable 5 偏向柔和的光球，而 Opus 4.8 创造了一个最繁忙、最像网络的粒子场，物理效果很棒，但视觉糖果稍少。获胜者：GPT-5.5，凭氛围取胜。

第三轮：可游玩的打砖块游戏

在画布上构建一个可游玩的打砖块/破砖游戏：挡板跟随鼠标移动，球击碎彩色砖块，配有分数和生命值。

亲自试玩：Grok 4.5 · GPT-5.5 · Opus 4.8 · Fable 5

势均力敌。四个模型都一次性生成了精美、可实际游玩的打砖块游戏，包含分数、生命值以及发光的挡板。Grok 4.5 和 GPT-5.5 更偏向霓虹街机风（当我们的脚本打球时，GPT 甚至打出了一些分数）。如果要斤斤计较，这四个都达到了可交付的质量。获胜者：全体。

数据收据：速度与成本

漂亮的演示是一回事；每个模型运行成本多少？我们用自有测试框架自行测量（三个固定提示，涵盖编码、推理和摘要，各重复三次，输出 token 限制为 400），通过应用使用的相同提供商路径进行。吞吐量是输出 token 除以总墙钟延迟，对每个提供商统一测量。

这正是 Grok 4.5 大显身手的地方。它在半秒内就给出了首 token，流式输出速度约 110 tokens/秒（几乎是这里其他模型的两倍），并且是其中成本最低的回复，恰好体现了 xAI 宣称的“单位时间和成本的智能”定位。它的墙钟中位数看似中游，是因为输出较长（每次回答生成的 token 最多），而且尾部延迟有波动（P95 约 9s）。GPT-5.5 在短答案上最快，Opus 4.8 处于均衡的中间位置，而 Fable 5 最慢且最贵，这是为占据智能排行榜顶端所付出的代价。

奖励轮：画点奇怪的东西

编码是一个维度；空间想象力是另一个。我们要求每个模型仅用一份手写 SVG（无光栅、无库）绘制一个故意离谱的场景：一个在月球上行走的宇航员背上骑着一匹马。角色反转，两个形象，一个文件。

Grok 4.5	GPT-5.5
Claude Opus 4.8	Claude Fable 5
---	---

原始 SVG 文件：Grok 4.5 · GPT-5.5 · Opus 4.8 · Fable 5

Fable 5 大放异彩：一匹戴着牛仔帽的马大喊着“驾，人类！”，而弯腰驼背的宇航员气喘吁吁地“嗬...嗬...”。这不仅是绘画，更是喜剧。GPT-5.5 以兴高采烈的马喊出“WHEE!”紧随其后，获得第二。Grok 4.5 干净利落地完成了要求，场景色彩丰富、清晰可辨。Opus 4.8 同样绘出了富有个性的马骑宇航员图，但其原始 SVG 带有一个重复属性，会干扰严格的 SVG 解析器（我们在上图中进行了宽松渲染），这是一个微小但确实存在的正确性问题。

结论

• Grok 4.5 是速度与性价比的怪兽。 它构建出了出色的打砖块游戏和流畅的重力模拟，流式输出速度大约是同行两倍，且运行成本最低。它唯一真正的失误是在最困难的状态管理任务上（首次尝试魔方为空，重试后修复）。如果你面对的是延迟和成本成倍累积的大批量代码生成工作，那它非常难以否定。参见 Grok 4.5 对比 GPT-5.5。
• Claude Opus 4.8 和 Fable 5 是最可靠的建设者。 它们是仅有两个首次尝试就完成了 3D 魔方任务的模型，而且 Fable 贡献了最有趣的 SVG。你需要在延迟和价格上为此买单，尤其是 Fable。参见 Grok 4.5 对比 Opus 4.8。
• GPT-5.5 是灵敏的风格家。 拥有最漂亮的重力模拟，最短答案速度最快，但在魔方上失了手。

诚实的标题：在全新发布当日，Grok 4.5 与市面上最好的模型正面交锋，并在速度和成本上直接胜出。想用自己的提示一分高下吗？这里的每个模型都可以通过 一个 TryAI 账号 按量付费运行。浏览模型 并开始你自己的构建对决。

自己试试看

这里提到的所有模型均可在 TryAI 上用一个账号按量付费使用，无需订阅。

免费开始

Grok 4.5 landed this week with xAI calling it their smartest model yet, "trained alongside Cursor" for coding and agentic work. The internet's favorite way to vibe-check a new coding model is the build-off: hand it one prompt, see what app it spits out, post the clip. So we did it, but properly.

We gave Grok 4.5, GPT-5.5, Claude Opus 4.8, and Claude Fable 5 the exact same three prompts: build a single self-contained HTML file (no libraries, no network calls) for an interactive app. One shot each, no hand-holding, no prompt-fiddling. Then we loaded every result in a real browser, poked at it, and recorded what happened. Every clip below is the model's raw output, and you can click "play it live" to run the actual generated app yourself.

One rule on failures: if an app did not render at all, we allowed exactly one retry, and we tell you when we used it.

Round 1: a 3D Rubik's Cube that scrambles and solves

Build a colorful, 3D-looking Rubik's Cube with 'Scramble' and 'Solve' buttons. The cube must visibly animate its rotations.

This is the mean one. It needs real 3D math, per-face state, and animation, all in one file.

|

Play them live: Grok 4.5 · GPT-5.5 · Opus 4.8 · Fable 5

The Claudes ran away with this one. Opus 4.8 and Fable 5 both produced a proper, correctly colored 3D cube that scrambles and solves with animated turns, first try. Grok 4.5 stumbled here: its first attempt rendered the title and buttons but no cube at all (a blank void), so we spent our one allowed retry, and the second attempt produced a clean, colorful 3D cube. GPT-5.5 was the odd one out, rendering only a single dark face with barely any color instead of a full cube. Winner: Opus and Fable, tied.

Round 2: a particle gravity sandbox

An interactive particle gravity sandbox on a canvas: hundreds of particles with trails, clicking adds a heavy attractor. Make it mesmerizing.

|

Play them live: Grok 4.5 · GPT-5.5 · Opus 4.8 · Fable 5

Everyone shipped a working sandbox here, so this came down to taste. GPT-5.5 made the most genuinely mesmerizing one: glowing neon attractors with dense, swirling colored trails. Grok 4.5 went clean and orbital, with tidy attractor rings and colorful streaking particles. Fable 5 leaned into soft glowing orbs, and Opus 4.8 produced the busiest, most web-like particle field, great physics, slightly less eye-candy. Winner: GPT-5.5, on vibes.

Round 3: a playable Breakout game

A playable Breakout / brick-breaker on a canvas: paddle follows the mouse, ball breaks colorful bricks, with score and lives.

|

Play them live: Grok 4.5 · GPT-5.5 · Opus 4.8 · Fable 5

Dead heat. All four produced a polished, actually-playable brick-breaker with score, lives, and a glowing paddle, on the first try. Grok 4.5 and GPT-5.5 leaned hardest into the neon arcade look (GPT even racked up a score while our script batted the ball around). If we are splitting hairs, all four are ship-quality. Winner: everybody.

The receipts: speed and cost

Pretty demos are one thing; what does each model cost to run? We measured it ourselves with our own harness (three fixed prompts spanning coding, reasoning, and summarization, three reps each, capped at 400 output tokens) through the same provider path the app uses. Throughput is output tokens over total wall-clock latency, measured uniformly for every provider.

This is where Grok 4.5 flexes. It hit first token in under half a second, streamed at ~110 tokens/second (roughly double everything else here), and was the cheapest reply of the bunch, exactly the "intelligence per unit of time and cost" pitch xAI made. Its median wall-clock only looks mid-pack because it is verbose (it wrote the most tokens per answer), and its tail was spiky (a ~9s p95). GPT-5.5 was snappiest on short answers, Opus 4.8 sat in the balanced middle, and Fable 5 was the slowest and priciest, the tax you pay for the top of the intelligence charts.

Bonus round: draw something weird

Coding is one axis; spatial imagination is another. We asked each model for a single hand-authored SVG (no raster, no libraries) of a deliberately silly scene: a horse riding piggyback on an astronaut walking on the moon. Role reversal, two figures, one file.

Grok 4.5	GPT-5.5
Claude Opus 4.8	Claude Fable 5
---	---

Raw SVGs: Grok 4.5 · GPT-5.5 · Opus 4.8 · Fable 5

Fable 5 stole the show: a cowboy-hatted horse yelling "Giddy-up, human!" while a hunched astronaut wheezes "hff... hff...". That is not just drawing, that is comedy. GPT-5.5 was a close second with a gleeful horse shouting "WHEE!". Grok 4.5 nailed the brief cleanly with a colorful, readable scene. Opus 4.8 drew a characterful horse-on-astronaut too, but its raw SVG shipped with a duplicate attribute that trips strict SVG parsers (we rendered it leniently for the image above), a small but real correctness ding.

The verdict

• Grok 4.5 is the speed-and-value monster. It built a great Breakout and a slick gravity sim, streams about twice as fast as the field, and is the cheapest to run. Its only real miss was the hardest stateful task (a blank cube on the first try, fixed on the retry). If your workload is high-volume codegen where latency and cost compound, it is very hard to argue with. See Grok 4.5 vs GPT-5.5.
• Claude Opus 4.8 and Fable 5 are the most reliable builders. They were the only two to nail the 3D cube first try, and Fable had the funniest SVG. You pay for it in latency and price, especially Fable. See Grok 4.5 vs Opus 4.8.
• GPT-5.5 is the snappy stylist. Prettiest gravity sim, fastest short answers, but it flubbed the cube.

The honest headline: on a brand-new launch day, Grok 4.5 went toe to toe with the best models you can buy and won on speed and cost outright. Want to settle it with your own prompts? Every model here runs on one TryAI account, pay-as-you-go. Browse the models and start your own build-off.

Try it yourself

Every model mentioned here is available on TryAI with one account, pay-as-you-go, no subscription.

Start free