Frontier Models Fail at Real Work

🔥 What happened Epoch tested six models on real work tasks – graphic design, research, data analysis. Claude Fable 5.1 and GPT-6 Astra lead, but fail on open-ended work. Open-weight models flop even on well-defined tasks. 💡 Why it matters Models miss Epoch's visual style, ignore implicit conventions, and produce overly dense outputs. They spot promising research directions but can't execute them. Existing benchmarks overstate progress because they only measure easily verifiable tasks. ⚡ Our take Anyone betting GPT-6 replaces entire teams should read this report. The gap between benchmark scores and real work is wider than the industry admits.