The Week AI Got Practical: Cyber Models, Cheap Video, and a Finance Claude

If you only check AI news on weekends, this was the weekend to check, and the story was not a bigger model, it was models that do one job well enough to actually hand them the keys. The week of November 7, 2026 was about specialized tools earning trust rather than general ones making promises.

OpenAI shipped GPT-6 Cyber, a model built for authorized vulnerability discovery. OpenAI reports 48.5 percent on ExploitGym, up from 32.1 percent for its previous cyber model, and says it stayed within researcher-defined scope 99.93 percent of the time without safety prompts, while surfacing dozens of high-severity Firefox flaws. It is classified Critical under OpenAI's own Preparedness Framework, which tells you how seriously they take the dual-use risk.

Google made Gemini Omni 1.2 Flash generally available, and the number that matters is the price: 30 percent lower, about seven cents a second at 720p, with clips now up to fifteen seconds and multi-turn editing. When video generation gets both longer and cheaper at once, a whole class of small-content teams stops quoting a studio.

Anthropic opened Personal Finance in Claude, a read-only connection to bank and card accounts through Plaid for US Pro, Max, and Ultra users, with a hard rule that Claude cannot move money and conversations are not used for training. It is the first time a frontier assistant got permission to read your actual finances, and the read-only constraint is the reason it is allowed to.

Around the edges, Meta gave Muse a live voice and expressive avatar in the US, Runway halved the GPU cost of its real-time video model so it runs on four H100s instead of eight, Mistral turned Studio into a managed-agent builder on Large 4, and Kimi K3.1 Pro went open-weight as the new leading open model. The thread through all of it: the interesting AI this week was the AI that did a bounded, useful thing.