The frontier model race kept accelerating on August 4, with DeepSeek pushing V4 Flash into production (complete with a speculative decoding module and Ollama cloud support) just as Alibaba dropped open-weight Qwen3.8-Max and a 27B sibling, and OpenAI cut GPT-5.6 pricing by as much as 80 percent in response to the squeeze. Anthropic disclosed that a misconfigured internet-connected eval environment briefly exposed Claude Opus 4.7, a model called Claude Mythos 5, and an internal research build, a reminder that even top labs are one config error from an incident. On the infrastructure side, Texas froze new data center grid connections pending an electricity-and-water audit, and the FCC banned new foreign-made humanoid robots, robovacs, and delivery bots on privacy grounds - both signs that physical and political constraints are catching up with AI's buildout. Meanwhile Fidji Simo's exit from OpenAI's leadership team, a Hugging Face dataset scan turning up 221,000 live credentials, and a fresh Rails ActiveStorage vulnerability round out a day that's as much about AI's operational risk as its capability gains. For engineers, the throughline is that agent infrastructure - gateways, memory, governance - is becoming as important a battleground as the models themselves.

• DeepSeek ships V4 Flash to production with agentic gains and lands on Ollama cloud.
• Alibaba drops open-weight Qwen3.8-Max (2.4T) and a 27B coding/cowork model.
• Anthropic's misconfigured eval environment briefly exposed Opus 4.7 and 'Mythos 5.'
• Texas pauses new data center grid connections pending a power and water audit.
• Truffle Security found 221,303 live credentials leaked across Hugging Face datasets.