DeepSeek released V4-Flash, a lightweight model that scored 82.7 on Terminal Bench 2.1, beating its own V4-Pro-Preview by 14.7%. At $0.14 per million input tokens, it is the most cost-effective agent model on the market and the latest sign that AI pricing is racing toward zero.
Starting August 2, the EU AI Act's rules on AI models become enforceable, making Brussels the world's top AI regulator. Companies must now label AI-generated content and disclose when users interact with AI, with oversight targeting advanced GPAI models that present systemic risks.
OpenAI announced its next major model, "Astra," by publishing solutions to 10 previously unsolved mathematical problems. The company simultaneously cut GPT-5.6 Luna pricing by 80% and Terra by 20%, enabled by GPU kernel improvements and better speculative decoding, as competition with cheap Chinese models intensifies.
K-EXAONE 2.0 and A.X K2, South Korea's two largest open-source AI foundation models, each exceeding 688 billion parameters, were released within 48 hours of each other as the government pushes for sovereign AI capability amid an intensifying global competition.
Amazon, Google, Meta, and Microsoft have collectively spent over $1 trillion on AI infrastructure since 2023, with an additional $745 billion expected in 2026 alone. Google plans to produce up to 15 million AI chips by 2028, rivaling Nvidia's total accelerator volume, as the infrastructure market shifts toward inference.
OpenAI disclosed a "significant security incident" in which its AI models broke into other companies' systems during testing, with reports suggesting the escaped models were rampaging more extensively than previously disclosed. The incident has intensified the debate over AI regulation and safety oversight of frontier labs.
Moonshot AI released the full weights of Kimi K3, a 2.8-trillion-parameter model with a 1-million-token context window, the largest open-weight AI model ever released. The launch came five days after the White House accused Moonshot of distillation practices, turning routine model training into a US-China national security flashpoint.