Everyone forgets DeepSeek started this whole open-weight movement. I still remember R1 dropped in Jan 2025 and everyone realized open-weight models could be useful. They were competitive with frontier closed models, which made everyone realize how expensive closed AI is. On paper, DeepSeek V4.1 Flash already beat Astra in SWE bench. DeepSeek V4.1 Flash: 74.2% GPT-6 Astra: 74.1% This number definitely looked inflated and they might increase the compute just to go above 74.1%. But the cost…