一夜之间，GPT-6.2 Astra 和 Fable 5.5 双双曝光，AI 圈疯狂 11 月要来了
刚进入 11 月，AI 圈又开始热闹起来了。
过去十几个小时里，Anthropic 和 OpenAI 接连传出新模型消息。有人在 Anthropic 官网源代码中发现了 Claude Mythos 5.5 的名字，还有开发者晒出疑似 Claude Fable 5.5 生成的 3D 游戏和交互作品。
另一边，OpenAI 尚未官宣的 GPT-6.2 Astra，也出现在部分用户的 Codex 使用统计和 API 响应记录中，甚至有人声称已经成功调用。
两家公司的新模型线索几乎同时出现，结合此前关于 Qwen4、GLM-5.5 和 Hy4 的传闻，也让 AI 社区开始猜测，11 月是否又会迎来一轮旗舰模型集中更新。
Claude 5.5 官网意外曝光，新模型要来了？
事情要从 Anthropic 官网的一次意外发现说起。
11 月 1 日深夜，科技博主 @kimmonismus 在 X 平台发帖称，Anthropic 官方网站的 HTML 源代码中，出现了尚未正式公布的 Claude Mythos 5.5。
根据截图，Anthropic 的 Mythos 产品页面中，隐藏的 H1 标题以及图片的 alt 属性，都包含 Mythos 5.5 字样。
随后，科技媒体 TestingCatalog 也关注到这一变化，并推测 Anthropic 可能正在筹备新一代模型。
对于不熟悉 Anthropic 产品线的读者，可以简单理解为，Mythos 主要面向网络安全和生物学等高风险研究领域，访问权限受到严格限制；Fable 则在相同底层模型基础上加入安全防护，面向更广泛的用户提供服务。
因此，Mythos 5.5 的名称意外现身官网，也让外界开始猜测，面向普通用户的 Fable 5.5 是否即将登场。
值得注意的是，今天凌晨，开发者 @chetaslua 进一步发现，Mythos 5.5 曾在官网 HTML 中出现六次，但正文介绍仍然停留在 Mythos 5.1。
向左滑动查看更多内容
随后，另一位用户 @pankajkumar_dev 表示，Anthropic 已经删除了涉及 Mythos 5.5 的网页内容。
与此同时，另一条关于 Fable 5.5 的线索也被翻了出来。
早在 10 月 13 日，开源大模型 API 网关 LiteLLM 的 Azure 价格配置中，就曾出现过 azure_ai/claude-fable-5-5 的模型标识。相关配置随后被移除，但第三方 MLflow 模型目录仍保留了对应记录。
根据流传的配置信息，Fable 5.5 可能支持 100 万 token 输入上下文、12.8 万 token 最大输出，每百万输入 token 为 10 美元，输出为 50 美元，缓存读取价格则为 0.25 美元。
至于新模型究竟什么时候发布，社区同样众说纷纭，有人预测本周，有人预测下周，也有人认为要等到 11 月中旬，因为这可能是 Anthropic 上市的最佳献礼。
Claude Fable 5.5 编程能力曝光，3D 游戏都能一句话生成？
相比官网源代码里的几个字符，更让开发者兴奋的，是 X 平台上接连出现的 Fable 5.5 演示视频。
11 月 1 日，开发者 @notjazii 发布了一段约 36 秒的对比视频，声称使用相同提示词，让 Claude Opus 5.5 和 Fable 5.5 分别生成 3D 场景。
画面中，Opus 5.5 主要呈现湖泊、公园、喷泉以及四季变化的自然景观，整体更接近微缩模型展示。而 Fable 5.5 一侧则出现了农场四季变化、城市建筑逐渐生长、港湾昼夜交替，甚至还有火山喷发、火箭升空等动态事件。
相比静态场景，Fable 5.5 的作品更强调连续动画、时间变化和物体状态。
发布者认为，新模型在此次对比中全面胜出。
紧接着，他又展示了一款据称由 Fable 5.5 生成的 Minecraft 风格游戏。
在约 92 秒的视频中，玩家能够以第一人称视角探索方块世界，穿过森林、洞穴、海岸和沙漠河流，甚至还能看到熔岩区域、粉色树冠森林，以及水面上的动态光照。
游戏界面同样相当完整，包括镐子、物品快捷栏和分类背包等常见元素。作者声称，整个作品只用了一个提示词，花费一到两个小时便完成生成。
另外两段演示也颇有看点。
开发者 @abhinavflac 展示了一艘据称由 Fable 5.5 通过 zero-shot 提示词生成的交互式历史帆船。
镜头能够围绕船身移动，观察甲板、炮门、帆面和绳索结构，海浪、船尾尾流以及夕阳下的水面反光也被呈现出来。
另一位开发者 @chetaslua 则展示了一款基于 Three.js 的跑车应用，用户可以从不同角度观察车辆，还能随时切换红、蓝、银、黑和荧光绿等车漆颜色。
从这些作品来看，开发者已经开始尝试让 AI 承担更复杂的 3D 建模、场景动画和交互开发任务。不过，相关视频并未附带完整提示词、调用日志和源代码，模型身份、生成耗时以及是否一次完成，仍主要依赖作者自述。
倘若后续能够得到验证，Fable 5.5 在复杂交互应用开发方面的表现，确实值得期待。
GPT-6.2 Astra 也来了，Codex 用户提前用上新模型？
Anthropic 这边消息不断，OpenAI 也出现了新的动静。
11 月 1 日，X 用户 @TokenGremlin 发帖称，部分用户的 Codex 使用统计中，出现了一个尚未正式公布的模型名称。
gpt-6.2-astra
根据网友的截图，在某个账户的模型使用占比中，GPT-6 Astra 占据约 80%，GPT-6.2 Astra 则占据 19.8%。
有意思的是，用户在模型选择界面看到的仍然是 GPT-6 Astra，但后台统计却显示了新的模型标识。
随后，开发者 @Mr_Salio 也晒出一段 API 响应数据，其中 response.model 字段显示为 gpt-6.2-astra。
两条线索分别来自使用统计和响应记录，让外界猜测 OpenAI 可能已经开始测试新版本 Astra。
更进一步，今天凌晨，用户 @ai_funss 表示，升级 Codex CLI 后，可以通过终端指定模型名称，尝试调用 GPT-6.2 Astra。
其展示的截图不仅出现了 gpt-6.2-astra medium 标识，还显示 1M window，也就是 100 万 token 上下文窗口。
至于性能，社区内已经出现两种截然不同的评价。
@TokenGremlin 认为，GPT-6.2 Astra 相比前代有明显进步，尤其是在设计、技术能力和编程表现方面，部分任务甚至可能超过 Claude Opus 5.5。
但另一位用户 @Da7_Tech 却认为，新 Astra 尚未达到 Opus 5.5 的水平，更不用说传闻中的 Fable 5.5。
值得一提的是，10 月 31 日，OpenAI 官方账号还发布了一段视频，配文「未来有很多种可能。你想创造哪一种？」，也引发了外界对其是否由新一代模型生成的猜测。
回顾最近一个月，全球大模型厂商的更新节奏明显加快。
10 月 28 日，Google DeepMind 刚刚发布 Gemini 4 Carbon，如今 Anthropic 和 OpenAI 又相继传出下一代模型的线索，国内的 Qwen、GLM、Hy4 等模型也被寄予新版本发布的期待。
如今的大模型竞争，已经逐渐进入一个有些荒诞的阶段。
一款旗舰模型还没正式亮相，网友就已经通过官网源代码、后台记录和演示视频，找到了各种蛛丝马迹，甚至提前争论起谁才是下一任世界第一。至于这些线索究竟是厂商无意留下的痕迹，还是有意放出的预热信号，恐怕只有厂商自己清楚。
但流量本身无疑是一把双刃剑。毕竟，AI 厂商最难击败的对手，可能已经不再是彼此，而是被上一代模型不断抬高胃口的用户。
当真实产品终于登场，如何超越互联网提前为它塑造的想象，或许比击败竞争对手更加困难。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
One night, GPT-6.2 Astra and Fable 5.5 both exposed, AI Circle crazy, November coming.
Just into November, the AI Circle is starting to get hot again.
In the last ten hours, Anthropic and Open AI have been connected to new model messages. The name Claude Mythos 5.5 was found in the Anthropic official web source code, as well as 3D games and interactive works generated by the developers of suspected Claude Fable 5.5.
On the other hand, Open AI's undeclared GPT-6.2 Astra also appears in some users' Cordex use of statistics and API response records, even with claims of successful calls.
The new model trails of the two companies appear almost simultaneously and, combined with previous rumours about Qwen4, GLM-5.5 and Hy4, allow the AI community to begin to speculate whether there will be another round of flagship model updates in November.
Claude, 5.5.
The story begins with an accident on the Anthropic network.
Late in the night of November 1st, tech blogger @kimmonismus posted on platform X that Claude Mythos 5.5 appeared in the HTML source code of the Anthropic official website that had not been officially published.
According to the screenshot, Mythos product page of Anthropic contains Mythos 5.5 words in hidden H1 titles and att properties of pictures.
Then, the technology media TestingCatalog followed the change and speculated that Anthropic might be preparing a new generation of models.
For readers who are not familiar with the Anthropic product line, it is simply understood that Mythos is mainly oriented towards high-risk research areas such as cybersecurity and biology, and access to permission is severely restricted; Fable adds safety protection based on the same bottom model and provides services to a wider audience.
Thus, the unexpected appearance of Mythos 5.5 ' s name on the official web also gave rise to speculation as to whether Fable 5.5 for ordinary users was about to appear.
It is noteworthy that in the early hours of this morning, the developer @chetaslua further discovered that Mythos 5.5 had appeared six times in the official network HTML, but that the text was still in Mythos 5.1.
Slide left to see more content
Another user, @pankajkumar dev, then said that Anthropic had deleted the content of the web page on Mythos 5.5.
At the same time, another clue about Fable 5.5 was turned out.
As early as October 13, the model identification of azure ai/claude-fable-5 was present in the Azure price configuration of the Open LLM API gateway LitELM. The configuration was subsequently removed, but the third party MLFlow model directory maintains the corresponding records.
Fable 5.5 may support 1 million token input context, 128,000 token maximum output, $10 per million input token, $50 for output, and $0.25 for cache reading, according to distributed configuration information.
As to when the new model will be publish, the community is equally divided, with some predicting this week, others predicting next week, and others suggesting that it will wait until mid-November, because this is probably the best gift for Anthropic to be listed.
Claude Fable 5.5 is exposed and 3D games can be generated in one sentence?
More exciting for developers than the characters in the official web source code is the Fable 5.5 demonstration video that appears on the X platform consecutively.
On 1 November, the developer @notjazi publish gave a comparison of about 36 seconds, claiming to use the same prompt to generate 3D scenarios for Claude Opus 5.5 and Fable 5.5 respectively.
In this picture, Opus 5.5 presents mainly lakes, parks, fountains and four seasons of changing natural landscapes that are closer to micro-models. On the Fable 5.5 side, there have been four-season changes on the farm, the growth of urban buildings, the alternation of harbours, and even volcanic eruptions and rocket lifts.
Fable 5.5 places more emphasis on continuous animation, time change and object status than static scenes.
According to publishers, the new model was fully successful in this comparison.
Immediately thereafter, he showed another minecraft-style game allegedly generated by Fable 5.5.
In about 92 seconds of video, players are able to explore the world of squares in terms of first person, through forests, caves, coasts and desert rivers, and even to see lava areas, pink crown forests, and dynamic light on the water surface.
The game interface is also quite complete and includes common elements such as snails, object shortcuts and sorted backpacks. The author claims that the whole work was produced using only one prompt, which took one or two hours to complete.
The other two presentations were interesting.
The developer @abhinavflac shows an interactive historical sailboat allegedly generated by Fable 5.5 via zero-shot prompt.
shot was able to move around the hull, observe decks, cannon doors, sails and rope structures, and the waves, the tails of the ship and the reflection of the water under the sunset were also shown.
Another developer, @chetaslua, demonstrated a Three.js-based motor vehicle application that allows users to observe vehicles from different angles and to change the paint colours of red, blue, silver, black and fluorescent vehicles at any time.
From these works, developers have started trying to get AI to undertake more complex 3D modelling, scene animation and interactive development tasks. However, the video was not accompanied by complete prompt, call logs and source codes, model identities, time-consuming generation, and whether or not they were completed, and remained largely dependent on the author ' s own account.
Fable 5.5's performance in the development of complex interactive applications is indeed to be expected if follow-up can be validated.
GPT-6.2 Astra is here, too. Codex users use new models in advance?
There's been a lot of news on this side of Anthropic and new movement on Open AI.
On 1 November, X-user @TokenGremlin posted that some of the users' Codex statistics showed a model name that had not been officially published.
gpt-6.2-astra
According to a screenshot of netizens, GPT-6 Astra accounted for about 80 per cent of the use of models in a given account, while GPT-6.2 Astra accounted for 19.8 per cent.
Interestingly, users still see GPT-6 Astra in the model selection interface, but back-office statistics show new model identifiers.
Subsequently, the developer @Mr Sario also tans an API response data which shows gpt-6.2-astra.
Two clues came from the use of statistical and response records, which led to speculation that Open AI might have started testing a new version of Astra.
Further, in the early hours of this morning, the user @ai funss indicated that after upgrading the Codex CLI, you can use the terminal to assign model names and try to call GPT-6.2 Astra.
It displays not only gpt-6.2-astra medium symbols, but also 1M window, or 1 million token context window.
As for performance, two distinct assessments have emerged within the community.
@TokenGremlin argues that GPT-6.2 Astra has made significant progress compared to previous generations, especially in terms of design, technical capability and programming performance, and that some tasks may even exceed Claude Opus 5.5.
Another user, @Da7 Tech, believes that the new Astra has not yet reached the level of Opus 5.5, let alone the anecdotal Fable 5.5.
It is worth mentioning that on October 31st, the official account of Open AI also made a video of publish, with the text "There are many possibilities for the future." Which one do you want to create?" It also triggers speculation about whether it was created by a new generation of models.
Looking back at the latest month, the pace of renewal of the global LLM manufacturer has increased markedly.
On October 28th, Google DeepMind was just publish Gemini 4 Carbon, and now Anthropic and Open AI are relaying the clues of the next generation of models, and domestic models Qwen, GLM, Hy4 and others were sent to the new version of publish.
Today's LLM competition has gradually entered a somewhat absurd phase.
Before a flagship model was officially presented, netizens had found all sorts of clues, even arguing in advance who was next in the world’s first place, via the official network code, backstage records and demonstration videos. It is only the manufacturer who knows whether these clues are unintentional traces of the manufacturer, or whether they are deliberately preheating signals.
But the flow itself is undoubtedly a double-edged sword. After all, AI's most difficult opponent to defeat may no longer be one another, but rather a user that has been constantly promoted by the previous generation of models.
When the real product finally arrives, it may be more difficult to go beyond the imagination that the Internet has shaped it in advance than to defeat competitors.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.