OpenAI Falling Behind in Real-World Coding Tasks

I've been tracking the Arena WebDev leaderboard closely, and the gap is starting to show.

Here's the current situation: Sol (OpenAI's latest) is the only OpenAI model holding a spot in the top 10. Meanwhile, 4 open-weight Chinese models have cracked frontier-level performance — and one of them is so affordable you can run it for days on what Opus costs per single task.

This isn't just benchmark gaming. These models are showing up in real developer workflows, code generation tools, and local agent setups where cost and performance matter.

What's more concerning is the accessibility angle. Open-weight models are being adopted by smaller teams and individual devs who can't afford API bills. When the performance delta closes and the price drops to near-zero, the moat disappears fast.

I'm curious if others are seeing the same shift in their tooling. Are your local dev environments running LLaMA-based models now? Is OpenAI's edge evaporating outside controlled benchmarks?

The coding arena leaderboard doesn't lie — and right now, it's telling a story that doesn't include OpenAI dominating like it used to.