Summary
Theo of T3 Chat argues that the gap between open-weight and closed frontier AI models is widening and is unlikely to close the way it did during the DeepSeek R1 moment in January 2025. A longtime open source advocate, he explains why he has stopped betting on open-weight models for serious work, even as he continues to build T3 Code as an open source product.
The video walks through his early experience using DeepSeek in T3 Chat, then makes the case that recent open-weight releases are overstated by their benchmark results. He describes a case on his own Skate Bench, where one open model scored perfectly on the original questions and then failed every newly added one. He compares claims about models such as DeepSeek V4 Flash, GLM-5 and Gemini 3 Flash against Anthropic’s Opus 5.5, and argues that real-world performance falls well short of the published numbers.
He also covers the compute side of the argument: scaling laws still reward more GPUs, import restrictions limit Chinese labs, and funding pressure pushes open-weight developers to chase benchmark wins. Viewers interested in open versus closed models, benchmark reliability and the economics of frontier AI will find a pointed, opinionated perspective here.
📺 Source: Theo – t3․gg · Published November 22, 2026
🏷️ Format: Opinion Editorial







