为什么新模型发布，都要先做一个游戏？
终于等到了！《影之刃零》发售了！
这款从 2023 年首次亮相、经过多轮实机演示和跳票（划掉）的国产动作游戏，今天正式登陆 PC 和 PS5。
再过三个星期，另一场持续多年的等待也将结束，《GTA 6》定于 11 月 19 日发售
一款大作面世，人们早已习惯以「年」为单位计算时间。开发团队要搭建城市和关卡，设计战斗规则，让角色、武器、物理与画面在玩家按下每一个按钮时正确合作。
另一边，几乎每款新的 AI 模型发布后，互联网上都会迅速出现一批游戏——速度堪比 Steam 上新。
Claude Opus 5.5 发布时，Anthropic 特意展示了不同模型根据同一条提示词制作游戏的结果，并称 Opus 5.5 在画面和完成度上得分最高。OpenAI 介绍 GPT-6 Astra 的计算机操作能力，也把游戏开发放进重点演示，一位开发者用它制作了一段可以游玩的伦敦历史。
▲ 图片来源：soonlab.ai
Kimi K3.1 Pro 上线后，社区又让它与 Claude、GPT 和 DeepSeek 同场制作第一人称射击、农场、赛车与迷宫游戏，比较生成时间、价格和最终效果。最新推出的 Genie 4，官方用的是「可交互世界」「实时模拟」「可探索原型」，仔细一看——不还是很像游戏嘛。
▲ 图片来源：Google DeepMind
模型发布后的传统节目原本是跑分，现在，越来越多人选择让它做一款游戏。这些当然不是下一部《影之刃零》、《GTA 6》。不过有意思的是，当 AI 公司需要证明一个新模型究竟强在哪里，它们为什么越来越喜欢让模型先做一款游戏？
一只怪鸟，如何变成了模型测试
10 月 28 日，独立开发者 Simon Willison 给这种变化下了一个判断：2025 年 11 月，代码 Agent 跨过了「可以成为程序员日常工具」的门槛；到了 2026 年 10 月，模型又跨过了制作游戏的另一道门槛。
这种展示方法有一个简单的前身。
2024 年，Simon Willison 开始用同一句提示词测试不同模型：「生成一张鹈鹕骑自行车的 SVG」，而且是模型直接写出构成图像的 SVG 代码，不能只是给个图。
任务听起来近乎玩笑，鹈鹕是哪里冒出来的怪鸟？但实际做起来并不容易：自行车需要两个对齐的轮子、相连的车架、车把与踏板；鹈鹕不仅要被画出来，还要坐在车座上，把脚放到踏板附近，身体与自行车保持合理比例。代码只要漏掉一个标签，图像便可能无法渲染；坐标关系稍有偏差，鹈鹕就会悬在车外，车轮和车架各过各的生活。
早期模型画出的结果通常只是两个圆、一个三角形和一只勉强能认出的鸟。到了 2026 年，新模型已经会给鹈鹕补上羽毛、长喙、脚蹬和车链，甚至主动增加鱼篮、帽子与沿途风景，景色要随着动作有所变化。在 Simon 的博客上，因此有了各种各样的鹈鹕。
「鹈鹕测试」并不是严谨 benchmark，它只运行一次，没有统一控制推理档位、成本和工具，也不能证明哪一个模型综合能力最强，它流行的原因在于不需要解释。空间关系是否合理、代码有没有正确变成图像，一眼就能看见。
游戏延续了同一种逻辑，只是把静止的 SVG 变成了可以操作的系统。如果说鹈鹕测试的是模型能否把一组代码组织成连贯画面。游戏测试的是，当用户有所操作后，模型设计出来的框架和组件，还能不能继续运作。
游戏里藏着多少种能力
以近期常见的「单提示词制作 FPS」为例，玩家拿着武器移动，敌人从远处靠近，开枪后出现命中反馈，击败一波敌人后进入下一关。
▲ Claude of Duty 项目，使用Claude Opus 5 统筹多个 Agent，写出约 5.5 万行 Three.js 代码，模型、纹理、动画和声音均由代码实时生成的仿《使命召唤》游戏。图片来源：Github @Claude of Duty
看着只是十几秒之间的操作，模型实际需要同时搭起许多相互依赖的部分：它首先要建立三维场景和第一人称相机，让鼠标控制视角、键盘控制移动；还要生成武器、敌人与障碍物，并处理它们之间的距离和遮挡。扣下扳机后，程序需要计算射线或子弹轨迹，判断有没有命中敌人的碰撞区域，再扣除生命值、播放动画、更新得分。敌人还要知道怎样寻找玩家，关卡系统则要决定下一波在何时、何处出现。
这些功能单独写对了，游戏仍然可能没法玩：相机转动与角色方向没有同步，玩家会横着滑行；子弹击中模型却没有触发碰撞，枪便只是会发光的装饰；敌人、计分与关卡各自运行，却没有共享同一套状态，游戏就无法形成循环。
▲ GPT-5.2 根据「制作浏览器版《反恐精英》」的提示生成的可玩 FPS。图片来源：instantdb.com
因此，一款简单游戏同时检查代码生成、空间理解、视觉设计、状态管理、物理规则、交互反馈和错误修复。它还要求模型完成最后那一点常被 Agent 忽略的工作：把零散功能接起来，交付一个用户真的可以操作的成品。
这正是游戏比「写一个登录页面」更适合展示模型力的地方。网页可以靠一张漂亮截图遮住按钮失效，游戏却会在玩家第一次移动、射击或撞墙时暴露问题。代码是否运行只是第一关，输入是否改变世界、世界是否及时回应，才决定它是不是真-游戏。
为什么是浏览器游戏
新模型演示里的游戏，大多不是需要安装的完整客户端，而是由 HTML、JavaScript、Three.js 或 Phaser 搭成的浏览器作品。
▲ 由 Claude Chat 制作的 RPG 游戏 Eberwick。图片来源：simonwillison.net
这类游戏对模型格外友好。前端库、教学代码和常见玩法在公开互联网中大量存在，模型熟悉怎样调用它们；浏览器又已经内置图形、声音、输入与网络能力，一份代码便能同时容纳画面和玩法。作品完成后只需放上静态网页，别人点开链接即可运行，不必配置引擎和下载文件。
游戏本身也提供了高度成熟的语言，不管是对于模型，还是用户这一点都成立。看到第一人称镜头，模型知道鼠标通常控制方向；看到赛车，便可以调用加速、碰撞、计时和检查点；看到塔防，则能从射程、冷却、资源和波次开始组织系统。它不必从零发明「什么叫游戏」，而是把训练数据中积累的大量惯例重新拼装起来。
▲由 Claude Chat 制作的 RPG 游戏 Eberwick。图片来源：simonwillison.net
这让游戏成为一种非常高效的模型展示：技术上足够复杂，传播上又足够轻巧。几万行代码最终被折叠为一段人物奔跑、车辆转弯或子弹命中的视频。观众不需要理解代码，也能判断画面是否穿帮、操作是否流畅。
但这种优势同样会美化模型。熟悉的玩法拥有大量现成范式，浏览器原型也不必处理账号系统、多人并发、不同硬件、版本更新和数年维护。模型做出一款能跑的射击游戏，与一个团队交付真正的商业游戏，仍然是两种工作。
从「像游戏」到「真能玩」
最近的变化并不只是模型更会写代码，Simon Willison 认为，Claude Opus 5.5 开始表现出一点过去模型缺少的游戏设计能力。
他用普通 Claude 聊天界面，只输入一次提示词，做出了《Tree Appreciator》。画面会并排生成两棵略有差异的三维树，玩家可以旋转和放大场景，找出鸟、松鼠或叶片等细小差别；完成之后，这棵树还会被收入一册虚拟标本集。
▲ 图片来源：simonwillison.net
从代码角度看，它仍然是一款很小的「找不同」游戏。真正让 Willison 意外的，是模型自行补齐了一套温和但完整的体验：如何告诉玩家目标，怎样逐渐增加难度，又如何用标本收集提供继续游玩的理由，组织出了开始、反馈与完成感。
另一个项目《Battle Drones》更能看出差异。早期版本由 Claude Fable 5 制作，已经拥有类似《命令与征服》的采集和战斗，但 Willison 觉得并不好玩。他只要求 Opus 5.5 「让它变得有趣得多」，后来又让模型将游戏迁移到 Three.js。现在，它拥有两至四人对战、七张地图、勋章、录像、随机地图，以及能够改变规则的 mutator。
▲ 由 Claude Chat 制作的 RTS 游戏 Eberwick，左为初代 Fable 5 制作，右边为 Opus 5.5 优化后版本。图片来源：simonwillison.net
模型对于「乐趣」的理解，未必像人一样，甚至不同的游戏玩家对这个词的理解也不尽相同。模型主要是识别出一款即时战略游戏，通常需要怎样的反馈密度、推进节奏和局势变化，再把这些惯例组合起来。但从展示角度看，评价标准已经发生变化：之前模型能不能把游戏做出来，做得出来已经属于是颇有能力。现在，得能实际玩下去。
不过，「一条提示词做出游戏」很容易制造一种错觉，仿佛模型在几分钟内完成了游戏公司多年来的开发心血。实际上，只从演示视频里，只能窥见一点点。
Claude Opus 5.5 制作的三维城市游戏《Neon Palms》，最初也来自一项游戏 benchmark。视频里，玩家可以在霓虹城市中驾驶和探索，所有画面都来自实际运行的程序，并非生成视频。
▲ 三维城市游戏 Neon Palms 实机演示
但它不是「一键直出」的结果，整个项目从 9 月 26 日持续到 28 日，Agent 分十一轮工作，共运行 69 次，最多六个 Agent 并行；按公开记录计算，模型处理了约 40.7 亿 Token，其中绝大部分是缓存上下文，输出约 320 万 Token。按 API 牌价估算，开发成本约为 1512 美元。展示视频还使用了之后加入小幅视觉修复和专门展示模式的版本，benchmark 本身仍记录着不少错误。
从绝对数字来看，这已经相当惊人，数天内完成可以运行的三维城市，成本还不高。不过仍然要记住，一段顺滑的 showcase 视频，只能展出结果，看不到的是提示词长度、重试次数、并行 Agent、人工选择和被毙掉的失败版本。
更系统的 GameCraft-Bench 也给出了相似结论，它要求代码 Agent 根据自然语言，在 Godot 中完成 140 个可游玩的游戏任务；评估不只检查代码，还会真正启动程序、回放操作并观看游戏表现。论文发布时，最强 Agent 的整体完成度仍只有 41.46%。模型很擅长迅速搭出让人惊艳的原型，却还不能稳定完成每一种规则、反馈和视觉要求。
▲ 图片来源：tongxuluo.github.io
模型「秀肌肉」的最佳方式
模型力内卷的下半场，任何新版本推出时都需要考虑，怎么给出一种大众能够看懂的证明方式。跑分、做题，都已经不足够直观，编程任务重，几十万行代码则很难浓缩易懂易传播的素材。
游戏却天然拥有画面、动作、声音与结果，它既是程序，也是成果；既能检查代码有没有运行，也允许观众亲自调试它。模型生成的每一个组件，都要在玩家按下按钮后与其他组件继续合作。
▲ 三维城市游戏 Neon Palms 场景图
如此一来，游戏已经成为比普通网页更密集、比编程项目更直观的能力展示舞台。新模型纷纷开始「做游戏」，并不代表它们已经掌握了完整的游戏工业，模型厂商也未必有这种想法。
它更像一种新的发布仪式，当模型的能力越来越难表现和说明，不如让它交出一个可以被观看、操作和传播的成品——除了游戏，又还能是别的什么呢。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Why does the new model publish have to start with a game?
Finally! Shadow's Blade Zero!
This is the national action game that officially landed on PC and PS5 today from the first appearance in 2023, through multi-wheeled demonstration and jumper.
In three weeks' time, another, many years' wait will be over.
It's been used to "years" as the time to calculate. The development team will build cities and gates, design rules of engagement, and allow role, weapon, physics and image to work together when the player presses each button.
On the other side, after almost every new AI model publish, there's a fast set of games on the Internet -- at a faster rate than Steam.
When Claude Opus 5.5 publish, Anthropic intentionally demonstrated the results of different models making games based on the same prompt and stated that Opus 5.5 scored the highest on the image and completion. Open AI describes GPT-6 Astra's computer skills and places the game in a focused presentation, where an developer produces a history of London that can be played.
Photo by Soonlab.ai
After Kimi K3.1 Pro went online, the community made it co-produced with Claude, GPT and DeepSeek the first person shooting, farm, car racing and maze games, comparing the time, price and end effect. The recently launched Genee 4, the official "real-time simulation" of the "interactable world" "explorable prototype", looks at -- it's still like a game.
Photo by Google DeepMind
The traditional programme after the model publish was originally a run, and now more and more people are choosing to make it a game. These are not the next ones, of course, The Shadows Zero, GTA 6. But what's interesting is, when AI needs to prove what a new model is, why do they increasingly like to let models play a game first?
A strange bird. How did it become a model test?
On October 28, the independent developer Simon Willison judged this change: in November 2025, the code Agent crossed the threshold of being a daily tool for programmers; in October 2026, the model crossed another threshold for making games.
This presentation has a simple front.
In 2024, Simon Willison started testing different models with the same sentence prompt: "to generate a cycling SVG for pelicans" and to write directly the SVG code that makes images, which cannot just give a picture.
The mission sounds like a joke. Where did pelicans come from? But it's not easy to actually do it: bicycles require two pairs of wheels, a connected frame, a car handle and a pedal; pelicans are not only painted, but also sit on the seat, with their feet near the pedal, with a reasonable ratio between the body and the bicycle. If one label is missing, the image may not be rendered; if the coordinates are slightly different, the pelicans will hang out of the car and each wheel and frame will live separately.
Early model results are usually two circles, a triangle and a bird barely recognized. By 2026, the new model would have filled pelicans with feathers, beaks, ankles and car chains, and even volunteered to add fish baskets, hats and viewing along the road, which would have changed as it moved. On Simon's blog, there's so many pelicans.
The pelican test is not a rigorous benchmark, it operates only once, without uniform control of the reasoning slotting, costs and tools, and does not prove which model is the most integrated, and it is popular because it does not need to be explained. It can be seen at first sight whether space relations are reasonable and whether the code is correctly translated into images.
The game continued the same logic, just turning static SVG into an operational system. If the pelican test is a model that can organize a set of codes into a consistent picture. The game test is whether the framework and components designed by the model can continue to function when the user is operational.
How many powers are hidden in the game?
In the case of the most common recent "single prompt FPS" player, who moved with a weapon, the enemy approached from a distance, fired a shot and received a shot back, and beat a wave of enemies and entered the next level.
▲Claude of Duty, using Claude Opus 5 to combine multiple Agents, to write about 55,000 lines Three.js code, a simulation of the Call to Mission game where models, textures, animations and sound are generated in real time by the code. Photo by Github @Claude of Duty
Looking at only a dozen seconds of operation, models actually need to combine many interdependent components: It starts with the creation of a three-dimensional scene and a single-name camera, allowing mouse to control the perspective, keyboard to control the movement, and the generation of weapons, enemies and obstacles, as well as the handling of distance and covertness between them. When the trigger is pulled, the procedure involves the calculation of a ray or bullet trajectory, the determination of the impact area of the enemy, the deduction of life value, the play of animations and the updating of points. The enemy also needs to know how to find a player, while the level system decides when and where the next wave will appear.
These functions are written in their own right, and the game may still not be able to play: the camera is not synchronized with the role direction, and the player slides across; the bullet hits the model without triggering a collision, the gun is just a glorified decoration; the enemy, the score and the level are running separately without sharing the same state, and the game cannot be revolving.
▲ GPT-5.2 Playable FPS based on the tips of "Producing a Browser Edition of Anti-Terrorism Elite". Image by instantdb.com
Therefore, a simple game examines both code generation, spatial understanding, visual design, state management, physical rules, interactive feedback and error repair. It also requires the model to do the last thing that Agent often ignores: to connect the fragmentation and deliver a product that the user can really operate.
This is where the game is better suited to show model power than writing a login page. The web page can override the button with a nice screenshot, while the game is exposed when the player first moves, shoots or hits a wall. Whether the code is running is only the first level, whether the input changes the world and whether the world responds in a timely manner determines whether it is real-play.
Why the browser game?
The game in the new model presentation is mostly not a complete client that needs to be installed, but a browser made up of HTML, JavaScript, Three.js or Phaser.
The RPG game by Claude Chat. Photo by Simonwillison.net
Such games are particularly friendly to models. Front-end libraries, teaching codes and common games are available on the open Internet, and models are familiar with how to use them; browsers are also built into graphics, sound, input and network capabilities, so that one code can accommodate both images and play. Once the work is completed, it can be run only on a static web page, and others can click on the link without having to configure the engine and download the file.
The game itself provides a highly mature language, both for models and for users. When the first person is known as shot, the model knows the direction the mouse usually controls; when you see a race car, you can call the accelerator, collision, timing and inspection points; when you see a tower, you can start with range, cooling, resources and wavelengths. It does not have to reassemble the vast amount of practice accumulated in training data from zero.
RPG game by Claude Chat. Photo by Simonwillison.net
This makes the game a very efficient model presentation: it's technically complex and it's easy to communicate. Tens of thousands of lines of code were eventually folded into a video of a person running, a car turning or a bullet shot. Audiences do not need to understand the code, and they can judge whether the image is running through or running.
But it's also going to be a model. The familiar play has a large number of off-the-shelf models and the browser prototype does not have to process account systems, multiple distributions, different hardware, version updates and years of maintenance. The model makes a running shooting game, delivering real business games with a team, which remains two jobs.
From "like a game" to "good play."
Recent changes are not just that models write codes, Simon Willison thinks Claude Opus 5.5 is beginning to show a little bit of the game design skills that were lacking in past models.
He used a regular Claude chat interface, entered prompt only once and produced Tre Appreciator. The image will generate two slightly different three-dimensional trees in parallel, and the player can rotate and magnify the scene to identify the small differences of birds, squirrels or leaves; when completed, the tree will be included in a virtual sample collection.
Source: Simonwillison.net.
From a code point of view, it's still a very small "seeing different" game. What really happened to Willison was that the model had built up a mild but complete set of experiences: how to tell players how to target, how to make it more difficult over time, how to use specimen collection to provide a reason to continue playing, and organized the start, feedback and completion.
Another project, Battle Drones, is more visible. The early version, produced by Claude Fable 5, already has a collection and fight like Command and Conquerion, but Willison doesn't find it funny. He only asked Opus 5.5 to "make it much more interesting" and then the model moved the game to Three.js. Now it has two or four men fighting, seven maps, medals, videos, random maps, and mutator who can change the rules.
▲ RTS game made by Claude Chat Eberwick, left for the original Fable 5, right for Opus 5.5. Photo by Simonwillison.net
The model's understanding of "fun" is not necessarily the same as that of human beings, or even that of different game players. The model mainly identifies an instant strategy game, which usually requires feedback intensity, a boost to the rhythm and a change of situation, and then combines these practices. From the point of view of presentation, however, the evaluation criteria have changed: the previous model was able to make the game and was already capable of doing it. Now, we have to be able to play.
However, "a prompt makes a game" can easily create a illusion as if the model had completed the game company's years of development in minutes. In fact, it's only a little bit from the presentation.
Claude Opus 5.5 produced a three-dimensional city game, Neon Palms, originally also from a game, benchmark. In the video, the player can drive and explore in Neon City, all of which comes from the actual program, not from the video.
3D city game Neon Palms
However, it is not the result of "one key out straight" and the whole project lasted from 26 to 28 September, and Agent worked in 11 rounds, running 69 times, up to a maximum of six Agent parallels; in public records, the model processed approximately 4.07 billion Token, the vast majority of which was cacheZcontext, and exported about 3.2 million Token. Development costs are estimated at approximately $1512 at API. The presentation video also used later versions that included small visual fixes and specialized displays, and Benchmark itself still recorded a number of errors.
In absolute terms, this has become quite alarming, and the cost of completing three-dimensional cities within days is not high. Still, it is important to bear in mind that a smooth showcase video can only show the results, not seeing prompt length, retry frequency, parallel Agent, manual selection and dead failure.
The more systematic GameCraft-Bench has similar findings, which require Agent to perform 140 playable games in Godot according to natural language; the assessment will not only check the code, but also actually start the program, play back and watch the game performance. At the time of the paper publish, the strongest Agent was still only 41.46 per cent overall. Models are very good at quickly developing amazing prototypes, but they are not yet stable in meeting every rule, feedback and visual requirement.
Source: tongxuluo.github.io
The best way to model "show muscles."
The second half of the roll within the power of the model, any new version of which needs to be introduced, will need to consider how to give a form of proof that the public can understand. Running points and making questions is no longer intuitive enough, programming tasks are heavy, and hundreds of thousands of lines of code are difficult to concentrate on asset, which is easy to understand.
But the game has natural images, actions, sound and results, and it is both a program and a result; it can check whether the code is running or allowing the audience to debug it. Each component generated by the model continues to work with other components after the player presses the button.
Neon Palms scene Figure
As a result, the game has become a more intensive and visual display of capabilities than a regular web page. The fact that new models are beginning to "play" does not mean that they already have the full game industry, nor do model manufacturers necessarily have that idea.
It's more like a new publish ritual, and when the power of the model becomes more and more difficult to demonstrate and demonstrate, it can hand over a finished product that can be seen, operated and disseminated — what else can it be?
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.