At dawn today, Opus 5.2 was quietly released. Many developers discovered that the new model, Claude Opus 5.2, has started a gray-scale test in Claude Code.
Last night, developers on X found that when calling Opus 5 in Claude Code, its performance was completely different from the web version of Opus 5, as if it were a "dimensional attack." This is a common "trick" used by leading AI companies — routing. Developers found through packet capture and checking request status that although the front-end name remained unchanged, the underlying model slug had already been directed to Opus 5.2. In other words, it's highly likely that Anthropic skipped the 5.1 version directly.
Fast, accurate, and ruthless, with automatic "intense loops"
According to developer feedback, this upgrade can be summarized in three words: fast, accurate, and ruthless. First, the response speed has increased dramatically. Compared to Opus 5's slow pace, the generation speed has made a noticeable leap after being routed to 5.2; second, the output is clean and concise, with minimal unnecessary content, significantly improving the quality of code generation and long text processing.
The most important thing is that it has completely eliminated the "laziness" problem. When facing complex long tasks, the current AI often likes to ask users to "continue" or just provide a framework for users to fill in themselves. However, Opus 5.2 is not lazy at all and is very "workaholic." A developer said that it doesn't even need to be prompted, and it will automatically enter an "intense loop," continuously self-iterate and improve the code until the task is completed.
How to determine whether you have been selected in the gray-scale test? Netizens discovered a clever "probe": ask "Do you know who 'Tibo the Resetter' is?" without connecting to the internet. The old training data and weights don't have this information. The web version of Opus 5 would be confused, while accounts routed to Opus 5.2 can accurately identify and explain the origin — inconsistent answers indicate different weights.
There are more hidden cards inside: Model 2 and RSI
As early as mid-August, a leaked internal risk report from Anthropic showed that Opus 5.2 was not the final card. According to the report, Anthropic has three plans to deal with OpenAI's GPT-6 Astra, and they may be combined.
The first layer is Claude Fable 5.2, which can be launched the fastest. Combined with the gray-scale test of Opus 5.2, the 5.2 family is expected to be fully launched in the market by the end of this month at the earliest, and by the end of next month at the latest. The second layer is the mysterious model internally called "Model 2," which scored 62.8% on the CoBench v2 list testing real AI research tasks, 12.5 points higher than the currently recognized strongest model Mythos 5. Internal sources said that most of the company's code is now written by Model 2 running Agent, and the official statement is that it is temporarily not being released externally. The third layer is RSI — recursive self-improvement. Official data states that the RSI model will replace up to 85% of the work done by its research team, with performance 22 points higher than Model 2.
