Many have attributed the recent wave of vibe-coded decompilations and recompilations of games to Claude Opus 5.5, Anthropic’s latest AI model that was released on September 22. Today, users are suddenly reporting that Opus 5.5’s ability to “reverse engineer anything” may have been stealthily nerfed.

To clarify, Anthropic hasn’t announced or released a new version of Opus 5.5. However, the company has revealed a string of updates across several Claude-related products on October 21. This lines up with the recent October 21 report of Opus 5.5’s nerfed reverse-engineering capabilities on r/Claude.

According to several users in the thread, prompting Claude Opus 5.5 to reverse-engineer any intellectual property will “trigger a safety guardrail mid sentence.” In one example, Claude stated that a “safety check stopped [its] response partway through” and that it could no longer “produce that write-up again, even reworded.”

Anthropic secretly killed the REA train (Reverse engineer anything)
byu/Clair_Personality inclaude

However, several users in the thread have disputed this. One user noted that asking Claude to reverse-engineer something was always a crapshoot, as “the classifier is just really twitchy.” Still, those disputing the potential change appear to be in the overwhelming minority.

But would Anthropic change something without stating said change in Claude’s patch notes? Yeah, apparently Anthropic does this all the time, so there is a solid chance that Opus 5.5’s reverse-engineering capabilities have silently been changed without informing users.

As for why Anthropic would choose to make its AI worse, well–I think I can wager a pretty good guess on that front.

Source link