Within eight days in early September 2026, DeepSeek-V4.1-Flash and Gemini 3.8 Flash toppled previous-generation $90/M-token flagship models on DeepSWE v1.1. Here is the forensic engineering breakdown: 890-byte KV cache, CED topology, RLVR benchmaxxing, and real-world agent TCO.