新智元报道
AI圈再次变天了！
Artificial Analysis刚更新了Intelligence Index，正式推出v5版，Claude Opus 5.5一举登顶，将OpenAI的GPT-6 Astra挤下神坛。
但没高兴多久，Gemini 4 Carbon发布，超越 Claude Opus 5.5夺魁！
但代号「Obsidian 3」的神秘模型意外浮出水面，而所有的线索都指向同一个名字：Anthropic尚未发布的 Fable 5.5。
AI榜单洗牌了。
Artificial Analysis发布了最新的 Intelligence Index v5 榜单。这是目前业内含金量最高、最难啃的测试之一。
结果出人意料，但又在情理之中。
Anthropic的 Claude Opus 5.5 以58分险胜，斩获第一。
OpenAI的 GPT-6 Astra 以57分，屈居第二。
戏剧性的是，谷歌随之发布了Gemini 4 Carbon，得分 AA 综合得分为59。
一分之差，AI 王座一夜之间两度易主。
但在编程智能体指数上，Anthropic仍以62 分领跑。
为什么说这次Anthropic登顶编程智能体指数榜含金量极高？因为测试规则变了，变得极其苛刻。
私有数据权重飙升至60%。
过去的AI测试，大家都在刷榜。模型把公开的GitHub代码库背得滚瓜烂熟，分数自然高。
这次不行了。官方直接把私有数据集的权重拉到了60%。
没有公开资料，没有历史作业。模型必须当场思考，当场解题。
靠「死记硬背」混日子的时代，结束了。
免疫学教授Derya Unutmaz评论称，新版本更合理，此前开源模型存在针对公开基准的过度优（benchmaxxing），与OpenAI、Anthropic及Gemini前沿模型的能力差距被清晰拉开。
不过，这并不意味着两者差距扩大了，而是衡量方式变化了。
引入高难度科研工作流。
新增了代码与科研工作流等地狱级挑战。
这不再是写个贪吃蛇代码那么简单，而是要求AI像真实的程序员和研究员一样，处理复杂的工程逻辑。
在这样的高压测试下，Opus 5.5 扛住了。它证明了Anthropic在复杂逻辑推理和真实场景落地上的恐怖实力。
Opus 5.5 登顶，只是前菜。
真正的硬核大瓜，藏在另一家机构的数据库里。
昨天，知名AI测试机构 Cognition 发布了其私有编程基准的最新版本：FrontierCode 2.0。
这是一个极其变态的测试：包含132项真实开源代码库中的任务。
评分标准简单粗暴：代码提交后，项目的维护者是否愿意合并（Merge）你的更改。
只要有一项阻断性检查未通过，直接计0分，没有任何同情分。为了防止数据污染，任务全程保密。
而且Artificial Analysis还要在 Stirrup 框架中每项任务跑三次，杜绝偶然性。
在这个榜单上，Claude Opus 5.5 (max) 以 46% 的成绩领跑全场。
GPT-6 Astra (max)、GPT-6.1 Terra (max) 和 Claude Sonnet 5.5 (max) 紧随其后，全部死磕在 42%。
但有人扒开了官网的数据接口。在 data.json 的隐藏代码里，发现了一个隐藏模型Obsidian 3。
数据令人倒吸一口凉气：
Obsidian 3，高推理模式得分 46.8%。
它硬生生把当前第一的 Opus 5.5（45.5%）踩在了脚下。
这个隐藏冠军，到底是谁？
社区的极客们迅速锁定了证据：测试框架。
Obsidian 3 运行在 claude-code 环境下。这是一个专属的测试框架，目前全世界只有六个 Claude 家族的模型在使用。
真相呼之欲出：这可能就是 Anthropic 尚未官宣的新一代大杀器——Fable 5.5。
争议却随之而来。
一部分人感到失望，直言这提升不够「性感」。
46.8% 对比 45.5%，仅仅提升了 1.3 个百分点。在最大（max）模式下，两者甚至持平，都是 45.5%。
更要命的是成本。
数据显示，在相同投入程度下，Obsidian 3 的成本约为 Opus 5.5 的 3.6 倍。在 high 模式下，单任务成本高达 8.62 美元；在某些极端的 max 模式下，单任务成本甚至飙升至约 47 美元。
跑一次任务，花掉三百多块人民币。这对于多数开发者来说，是难以承受的数字。
这种定价颇有一种 「AI 中的爱马仕」的奢华感。这倒也间接排除了主打性价比的Gemini。
「如果这就是 Fable 5.5，那看起来不够惊艳。」有开发者在 X 上抱怨。
而且你让它思考得越多，表现反而越差，感觉像是一个早期检查点，不太可能是最终的发布版。
这更像是一个 checkpoint。
但也有人看到了不同的东西：
不要被编程得分骗了，Fable 5.5 的核心定位，从来不是一个只会写代码的「码农」。
业内人士指出，Obsidian 3 虽然在编程成功率上只是小幅领先，但其在「中高推理区间」表现极其稳定。
它真正的跃升，在别的地方：
- 推理与直觉。
- 推断人类的真实意图。
- 提升任务的抱负水平。
- 创造性的「分布外」思维。
- 发明前所未有的新颖解决方案。
在复杂的真实科研流中，解决一个「无人碰过」的新问题，比修100个已知的Bug要有价值得多。
此外，当前泄露的只是测试节点数据，极有可能并非 Fable 5.5 的最终发布版本。Anthropic 还在进行价格与性能的最终调优。
如果 Cognition 已经对其进行了基准测试并不慎泄露，说明一切已经准备就绪。
可能明天，可能后天，AI界的又一次大地震，已经进入倒计时。
这是一个充满戏剧性的时间节点。
Opus 5.5 险胜 GPT-6 Astra，证明了 Anthropic 在编程上的控制力。
Gemini 4 Carbon反超Opus 5.5，证明谷歌宝刀未老，尚能一战。
而隐藏冠军Obsidian 3的暴雷，预示着 Fable 5.5 已经上膛，随时准备扣动扳机。
AI模型没有不败的神话，只有永远的内卷。
今天，Gemini 4 是第一。明天，Claude Fable 5.5 可能就会掀翻桌子。之后，OpenAI 的反击必将降临。
准备好吧，AI 模型发布的时间间隔可能会越来越短。
参考资料：
编辑：大卫
不喜欢
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。
New Ji-won.
The AI Circle has changed again!
Artificial Analysis has just updated Inteligence Index and officially launched the v5 edition, with Claude Opus 5.5 at the top, crowding down Open AI ' s GPS-6 Astra.
But not for long, Gemini 4 CarbonZ94Z, overcame Claude Opus 5.5!
But the mysterious model codenamed Obsidian 3 accidentally surfaced, and all the clues point to the same name: Anthropic's Fable 5.5 without publish.
AI list shuffled.
Artificial Analysis publish has the latest Intelligence Index v5 list. This is one of the most expensive and most difficult tests in the industry.
The result was unexpected, but reasonable.
Anthropic Claude Opus 5.5 wins by 58 points, first.
Open AI's GPT-6 Astra is in second place with 57 points.
Theatrically, Google followed publish to Gemini 4 Carbon, score AA 59.
The difference between one and the other is that the throne has changed twice a night.
But on the AI agent index, Anthropic is still running at 62.
Why do you say that this time the Anthropic top-programming AI agent index has a very high gold content? Because the test rules have changed and become extremely harsh.
Private data weights skyrocket to 60 per cent.
Past AI tests, everybody's on the list. The model decorates the open GitHub code library with natural high scores.
Not this time. Officially, the weight of private data sets has been brought directly to 60 per cent.
No public information, no historical work. Models have to be on the spot, on the spot.
It's over with the days of hard-on-death.
Professor Derya Unutmaz of Immunology commented that the new version was more rational, with the previous open-source model containing over-optimal (benchmaxing) for open benchmarks, and the capability gap with the forward models of Open AI, Anthropic and Gemini being clearly closed.
However, this does not mean that the gap between the two has widened, but rather that the measurement has changed.
workflow.
A new code has been added to the hell-class challenges such as scientific workflow.
This is no longer as simple as writing a snake code, but requires AI to handle complex engineering logic like real programmers and researchers.
Under such high pressure tests, Opus 5.5 is holding. It demonstrates the terrorist power of Anthropic in complex logic and real scenery.
Opus 5.5 tops, just the front course.
The real hard core is hidden in another agency's database.
Yesterday, the famous AI testing agency Cognition publish updated its private programming benchmark: FrontierCode 2.0.
This is a very sick test: it contains 132 real open source library tasks.
Ratings are simple and rough: if the project maintainer is willing to merge your changes after the code is submitted.
As long as a blockage check is not passed, the score is zero and no compassionate. In order to prevent data contamination, the mission is confidential.
Moreover, Artificial Analysis runs three times every mission within the Stirrup framework to stop coincidence.
On this list, Claude Opus 5.5 (max) leads the entire field at 46% of the grades.
GPT-6 Astra (max), GPT-6.1 Terra (max) and Claude Sonet 5.5 (max) followed and all died at 42 per cent.
But someone broke the data interface on the official network. In the hidden code of Data.json, a hidden model Obsidian 3 was found.
The data gives you a chill:
Obsidian 3, high reasoning mode score 46.8 per cent.
It's hard-boiled at the bottom of the current number one Opus 5.5 (45.5%).
Who is this hidden champion?
The community ' s extremists quickly locked out the evidence: a test framework.
Obsidian 3 runs in claude-code environment. This is an exclusive test framework, and there are only six Claude models in use around the world.
The truth came out: this could be the undeclared new-generation murder weapon of Anthropic - Fable 5.5.
The controversy followed.
Some people are disappointed that this is not "sexy" enough.
Compared to 45.5 per cent, the increase was only 1.3 percentage points. Under the largest (max) model, both are even, 45.5 per cent.
What's worse is the cost.
The data show that at the same level of input, the cost of Obsidian 3 is about 3.6 times that of Opus 5.5. In the high model, the cost of a single mission is as high as $8.62; in some extreme max models, the cost of a single mission even surges to about $47.
Run one mission and spend more than 300 yuan. This is an unbearable figure for most developers.
This pricing is quite a luxurious sense of "a horse in AI." This, in turn, indirectly excludes Gemini, who is the main price for money.
"If this is Fable 5.5, it doesn't look amazing. There are developers complaining on X.
And the more you think about it, the worse it is, the more you feel like an early check point, the less likely it is the final version of publish.
It's more like a checkpoint.
But someone saw something different:
Don't be fooled by programming scores, Fable 5.5's core location, never being a code-writing "coding farmer."
According to industry sources, Obsidian 3, although only a small lead in programming success rate, performed in an extremely stable manner in the "middle-high-level reasoning zone".
It's a real leap, somewhere else:
- Logic and intuition.
- Deduce the true intent of humanity.
- Raise the level of ambition of the mission.
- Creative "outside" thinking.
- To invent innovative solutions that are unprecedented.
In a complex and real scientific stream, solving a new "untouched" problem is much more valuable than repairing 100 known Bugs.
In addition, only test nodes data are currently leaked, most likely not the final version publish of Fable 5.5. Anthropic is still in the process of final adjustment of prices and performance.
If Cognition has already tested its benchmark, it is not careful to disclose that everything is ready.
Maybe tomorrow, maybe the day after tomorrow, another big earthquake in the AI world has entered countdown.
It's a dramatic time node.
Opus 5.5 Risk wins GPT-6 Astra, proving the control of Anthropic in programming.
Gemini 4 Carbon counter-opus 5.5, proving that Google is still alive and capable of war.
And hiding the thunderstorm of the champion Obsidian 3, premonitioning that Fable 5.5 is loaded and ready to pull the trigger.
The AI model has no undefeated myths, only the inner scrolls forever.
Today, Gemini 4 is number one. Tomorrow, Claude Fable 5.5 may turn over the table. After that, Open AI's counterattack will come.
Get ready. The AI model publish may be shorter and shorter.
References:
Editor: David
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.