[AINews] not much happened today

DeepSeek released V4-Flash 0731, a post-training-only update to its open-weight model that improves agent performance and coding benchmarks without changing the model’s architecture or parameter count (284B total, 13B active, 1M context). The update moves the model to near GPT‑5.6 Luna on the quality‑versus‑cost frontier while remaining cheaper—Artificial Analysis reports it lands at index score 50 vs. 51 for GPT‑5.6 Luna, with roughly 60% lower cost per task on DeepSeek’s API when the unusually aggressive 98% cache‑hit discount is applied. Because the gains come entirely from post‑training refinements for tool‑use and long‑horizon tasks, the release signals that frontier‑class capability can be advanced without pretraining scaling, reinforcing open‑weight competitiveness. The weights were released under MIT immediately alongside the API beta, and community integration into coding stacks (Codex, Cline, Hermes Agent) and local quantized deployment (lossless 4‑bit ~168 GB RAM) followed within hours. The release sharpens the price‑compression trend after OpenAI’s GPT‑5.6 cuts, and community discussion frames it as a practical win for open‑source defense arguments in current cyber‑saft