Anthropic's AI downgrade stings power users

Anthropic users across online forums are raising the same complaint: Claude suddenly feels ... bad. The backlash lands just as Anthropic is testing a more powerful model—Mythos—raising questions about whether cutting-edge AI is becoming less accessible even as it gets more capable. Over the past few weeks, users on X, GitHub, and Reddit have been swapping anecdotes, benchmarks, and prompts in an effort to pinpoint what changed and why. "Claude has regressed to the point it cannot be trusted to perform complex engineering," an AMD senior director wrote in a widely shared post on GitHub. Others have posted side-by-side outputs and benchmarks they say show Claude generating answers that are less accurate or nuanced. Much of the speculation centers on whether Claude has been deliberately scaled back—what users are calling "nerfed"—either to control costs or to redirect scarce compute toward Mythos and other frontier efforts. Anthropic says it adjusted the default level of reasoning in Claude Code, but denies the changes were tied to compute constraints or Mythos.