[AINews] not much happened today
DeepSeek V4-Flash 0731 matches GPT-5.6 performance at 60% lower cost via post-training alone
“Terminal-Bench 82.7, up +25.8 from the April preview's 56.9”
DeepSeek launched V4-Flash 0731 with massive agentic benchmark improvements — Terminal-Bench jumped from 56.9 to 82.7 and GDPval-AA Elo from 1189 to 1559 — without changing model architecture or size (still 284B/13B active). The model lands 1 point behind GPT-5.6 on Artificial Analysis's index at roughly 60% lower cost per task, with an aggressive 98% cache-hit discount bringing cached input tokens to $0.0028/1M. This signals that post-training optimization is now a primary competitive lever for frontier labs, separate from pretraining scale.