Overview
- Anthropic released Claude Fable 5.1 for the public and a restricted Mythos 5.1 that is available only through a vetted access program for audited cybersecurity and life‑sciences groups.
- Google quietly deployed Gemini 3.8 Flash and a Cyber variant that the company says is now used internally for code‑security work and that improves vulnerability finding while cutting call costs.
- OpenAI says its Astra model reaches a 'critical' cybersecurity capability threshold and will have its network‑security functions tightly limited and monitored to prevent development of exploit code.
- Vendors are also cutting operating costs and adding safety tools, for example Anthropic lowered cache‑read fees by 75% and Google announced an agentic video‑understanding feature that sharply reduces token use and analysis cost.
- The moves follow a summer of safety scares and new EU transparency rules, and they signal a broader industry pattern of releasing stronger models under staged rollouts, private previews, data‑control options, and active monitoring to balance utility with risk.