Anthropic dominates the July 2 news cycle from two directions: it shipped Claude Sonnet 5, a cheaper Sonnet tier tuned for agentic planning, tool use and coding, while the Commerce Department simultaneously lifted export controls that had blocked its Fable and Mythos models, letting Fable return to market. Google, Microsoft and Meituan all countered with their own model drops (Nano Banana 2 Lite, MAI-Code-1-Flash, and the 1.6T-parameter LongCat-2.0), underscoring how crowded and commoditized the model layer has become. On the infra side, Ollama squeezed a 90% Apple Silicon speedup out of Gemma 4 via multi-token prediction and PyTorch shipped a native RL post-training stack, while Moondream argued the GPU capacity buildout itself is overheated. Meanwhile security researchers are already probing the agentic wave for weaknesses, from fingerprinted China-linked API routers in Claude Code to a new zero-day exploit repo and a fresh Linux CVE - a reminder that shipping agents faster than you can secure them has a cost.

• Anthropic ships Claude Sonnet 5, a cheaper, more agentic Sonnet model
• US lifts export controls on Anthropic's Fable and Mythos models
• Google, Microsoft, Meituan all drop new models same week as Anthropic
• Ollama 0.31 makes Gemma 4 ~90% faster on Apple Silicon via multi-token prediction
• Researchers flag Claude Code fingerprinting China-linked API routers