Anthropic split its new model line into a public Claude Fable 5 and a gated Claude Mythos 5 reserved for vetted cyberdefenders, while Dario Amodei published a lengthy case for urgent AI policy given rising cyber and labor-market risk. Google countered with DiffusionGemma, a diffusion-based text model claiming 4x faster generation than standard Gemma, and Microsoft shipped MAI-Image-2.5 straight to the #2 spot on the image-editing arena. Money is following the models: Broadcom, Apollo, and Blackstone are standing up a $35 billion AI infrastructure platform for frontier labs. Meanwhile the tooling around agents is maturing fast, with Zscaler, Rain, and an open-source SafeAgentDB all tackling agent identity, spend, and data isolation, alongside a sharp critique of how well Claude Code actually reviews security-sensitive code. For engineers, the throughline is that model capability, access tiers, and the surrounding security/ops layer are all shifting simultaneously.

• Anthropic gates Claude Mythos 5 to select cyberdefenders while Fable 5 goes public
• Dario Amodei argues for urgent AI policy amid cyber and job-displacement risk
• Google's DiffusionGemma claims 4x faster text generation than standard Gemma
• Broadcom, Apollo, Blackstone launch $35B AI infrastructure platform for frontier labs
• Zscaler, Rain, and SafeAgentDB push agent identity, spend control, and data isolation