llm benchmarking

I ran the same tasks through ChatGPT, Claude, and Gemini in 2026. Here’s what broke, what surprised me, and which one I’d use for actual work tasks.