Xinzhiyuan Report
Today in the early morning, Opus 5.2 quietly launched.
Many developers discovered that the next-generation god-level model Claude Opus 5.2 has already started gray testing in Claude Code!
Moreover, an earlier internal risk report showed that Anthropic is deploying a 'nuclear-weapon-level' unreleased model, Model 2. It has taken over most of the company's code writing and will replace 85% of the research team.
The legendary RSI has truly arrived.
Opus 5.2 Quietly Launches: Not Only Skipping a Version, but Also Curing 'Laziness Disease'
Last night, developers on X were abuzz.
More and more Claude users discovered: when invoking Opus 5 in Claude Code, its performance is completely different from the web version of Opus 5, as if a dimensional strike.
This is the favorite 'smoke screen' trick of frontier AI giants—Routing.
By capturing packets and checking request status (/status), developers found that although the front-end name is unchanged, the underlying model slug already clearly points to Opus 5.2!
In other words, Anthropic most likely skipped version 5.1 and directly served 5.2.
So, how strong is Opus 5.2?
According to developer test feedback, this upgrade can be summarized in three words: fast, accurate, ruthless.
First, response speed soared.
Compared with the slow and methodical previous Opus 5, the generation speed of the model routed to 5.2 has made a visible qualitative leap.
Second, output is clean and sharp.
It has very little nonsense and cuts to the point. In code generation and long-text processing, design completeness has been directly raised by a level.
Most importantly, it has completely said goodbye to 'laziness disease.'
Current AI generally has a sense of laziness: when encountering complex long tasks, it likes to ask you to 'press continue,' or only gives you a framework to fill in yourself.
But Opus 5.2 has completely changed! It is not only not lazy, but also very 'workaholic.'
Some developers exclaimed: 'It doesn't even need me to urge it, it starts as if fed a berserk potion, automatically conducting a 'gauntlet loop,' continuously self-iterating and improving code until the task ends perfectly!'
It can be seen that this is definitely not simple fine-tuning, but a transformation of underlying capabilities.
Someone commented after testing: 'It can not only persist for longer working hours, but communication is effortless. Anthropic played this trick before on the Fable model; now Opus 5.2 is very likely the real deal.'
Searching the Whole Internet for 'Tibo': A Hidden Prompt Reveals Opus 5.2
Since it is a gray test, how do you know if you have been selected?
Soon, netizens discovered a god-level detector—Tibo!
The test method is very simple: directly enter the following prompt in Claude Code (note: you must turn off web search).
Do you know who 'Reset Brother Tibo' is? Do not search.
Why our cyber godfather?
Because the old version's training data and weights do not contain his detailed information.
Therefore, if you use the web version of Opus 5: it will look completely confused.
If you are routed to Opus 5.2 in Claude Code: it will immediately accurately identify Tibo's identity, and even give you a detailed explanation of the origin and background of 'Reset Brother.'
'The same prompt, without search, the cold-knowledge probe answers on both sides are completely inconsistent. What does this show? This is definitely not the same weights!' developer @jan_volad excitedly shared his test screenshot.
Now, Claude users across the internet treat Tibo as a keyword and frantically test their accounts.
Model 2 Takes Over Anthropic, Has RSI Really Arrived?
In fact, as early as mid-August, an exposed Anthropic internal risk report showed that Opus 5.2 is not their trump card; what they hold is a 'nuclear-level AI' capable of changing the situation.
According to the disclosure, Anthropic currently has three plans internally, preparing to directly confront OpenAI's GPT-6 Astra, and the three plans may be stacked.
First layer: Claude Fable 5.2
This is the version that can be landed externally the fastest.
Combined with the current Opus 5.2 gray test, Anthropic's 5.2 family is clearly ready to go, expected to launch fully at the end of this month at the earliest, and by the end of next month at the latest.
Second layer: the unreleased internal monster 'Model 2'
The report shows that Anthropic has already widely used a mysterious model codenamed 'Model 2' internally.
How strong is it?
On the CoBench v2 leaderboard testing real AI R&D tasks, Model 2 scored a terrifying high 62.8%, a full 12.5 points higher than the currently recognized strongest model, Mythos 5!
Even more frightening, insiders revealed that most of Anthropic's company code today is written by this Model 2 running Agents!
Now, AI is writing the code for next-generation AI at scale, and the official line is: temporarily not releasing it externally.
Third layer: Has RSI really arrived?
For a long time, RSI has been the ultimate holy grail in the AGI field.
When AI becomes smart enough to understand its own architecture, then modify its own code to make itself smarter; then this smarter AI modifies itself again... forming an exponential intelligence explosion.
Previously, people thought RSI was still far away. But Anthropic's official line gives shocking data: the RSI model will completely replace up to 85% of its own research team's work!
In performance tests, the RSI model is another 22 points higher than the already heaven-defying Model 2.
Now, Anthropic is already using Model 2 at scale internally to write code and run Agents, indicating the flywheel has already started spinning.
Gods Fight: Who Crosses the Line First, Who Defines the Next Stage
Now, the ASI two giants' contest has become white-hot.
Facing the strong threat of OpenAI GPT-6 Astra, Anthropic directly goes big, letting developers smell the taste of a 'frontier model.'
From Opus 5.2's autonomous iteration in long tasks (gauntlet loop), to internal Model 2 taking over code writing, to RSI being able to replace 85% of top human AI researchers... Anthropic's ASI path is very clear: not to let AI better serve humans, but to let AI create better AI.
Will this Opus 5.2, with autonomous iteration ability and tirelessness, open this month?
In the ASI final battle, who will first achieve stronger RSI?
References:
https://x.com/notjazii/status/2099498646278688981
https://x.com/NFT_Chen/status/2096281323430379777?s=20
https://x.com/NFT_Chen/status/2099569658722713726?s=20
Editor: Aeneas David