GPT-5.4 and Claude Opus Redefine AI Coding Limits

There’s a strange fracture in AI right now. The same technology that can autonomously rewrite an entire codebase over the course of an hour will, in its next breath, fumble a toddler’s question about Instagram reels. Andrej Karpathy calls it two groups talking past each other, and he’s right.

The free tier fumbles; the paid tier restructures, hunts vulnerabilities, and solves problems that demand rigorous verification. That gap isn’t a bug. It’s a signal.

When you can check an answer for correctness, code compiles or it doesn’t, a proof holds or it collapses, reinforcement learning with verifiable rewards drives breakneck progress. But in fuzzy domains like writing or consulting, where “right” is negotiable, gains stall. So here’s the real question bending researchers: can these models ever achieve general intelligence, or are they forever optimized for specific, checkable worlds?