Claude Opus 5.2 has quietly launched today and is currently in limited beta testing within Claude Code. Developer tests have revealed significant improvements in response speed and code generation quality, overcoming Opus 5’s previously sluggish performance. More importantly, the new version has overcome its “laziness issue”—no longer frequently prompting users to “continue” for complex, lengthy tasks, but instead proactively entering a “rigorous loop,” continuously iterating and refining the code until the task is fully completed. According to an internal Anthropic risk report that has been leaked, Opus 5.2 is not the final card; the company also holds a mysterious model codenamed “Model 2,” as well as the RSI system capable of replacing up to 85% of the research team’s workload. The goal is shifting from “making AI serve humans” to “making AI create better AI.”Author and source: AIBase
This morning, Opus 5.2 quietly launched. Many developers discovered that the next-generation model, Claude Opus 5.2, has begun gray testing in Claude Code.
Last night, developers on X discovered that when calling Opus 5 within Claude Code, its performance differed drastically from the web version of Opus 5, as if undergoing a downgrade. This is precisely the “smokescreen” commonly used by leading AI companies—routing. By inspecting network packets and request statuses, developers found that although the frontend name remained unchanged, the underlying model slug had already been redirected to Opus 5.2. In other words, Anthropic most likely skipped version 5.1 entirely.
Fast, accurate, and decisive, with an automatic "brutal cycle" feature.
Based on developer testing feedback, this upgrade can be summed up in three words: fast, accurate, and powerful. First, response speed has skyrocketed—unlike the sluggish performance of Opus 5, routing to 5.2 delivers a visibly dramatic improvement in generation speed. Second, the output is clean and concise, with minimal filler, raising the bar significantly in code generation and long-form text processing.
Most importantly, it has completely overcome the "laziness syndrome." While most AI systems today tend to prompt users to "click continue" or provide only a framework for users to fill in themselves, Opus 5.2 is not only non-lazy but also incredibly hardworking—some developers say it doesn’t even need prompting; it automatically enters a "gauntlet loop," continuously iterating and refining its code until the task is fully completed.
How can you tell if you’ve been selected by GrayScale? Netizens discovered a clever “probe”: ask, “Who is Reset Brother Tibo?” without enabling internet search. The older training data and weights contain no information on this, so the web version of Opus 5 will appear confused, while accounts routed to Opus 5.2 can accurately identify and explain its origin—the inconsistent answers indicate different weights.
There are still hidden cards inside: Model 2 and RSI
As early as mid-August, an leaked internal risk report from Anthropic revealed that Opus 5.2 was not their final card. It was disclosed that Anthropic has three internal strategies to counter OpenAI’s GPT-6 Astra, which may be combined.
The first layer is Claude Fable 5.2, which can be deployed externally the fastest; combined with the gray-scale testing of Opus 5.2, the 5.2 family is expected to be fully launched to market by the end of this month at the earliest, or no later than the end of next month. The second layer is the internally codenamed "Model 2," which scored 62.8% on the CoBench v2 leaderboard evaluating real AI development tasks—12.5 points higher than the currently acknowledged strongest model, Mythos 5; insiders claim that most of Anthropic’s internal code is now written by Model 2-powered agents, though the official stance is that it will not be released publicly for now. The third layer is RSI—Recursive Self-Improvement—official data states that RSI models will replace up to 85% of the work done by the company’s own research team and outperform Model 2 by an additional 22 points.
Faced with the strong threat of GPT-6 Astra, Anthropic’s path has become increasingly clear: from autonomous iteration in long tasks, to internal models taking over code writing, to RSI replacing most research work—the goal is shifting from “making AI better serve humans” to “making AI create better AI.”