Eleven Realtime is a family of speech-to-speech models where nothing in the stack is stitched together from someone else's parts. Recognition, voice activity detection, turn-taking, the language model and the voice are all picked and tuned by ElevenLabs as one system, and first audio comes back at 570ms on v1 Max and 620ms on Mini, across 90-odd languages. The API is OpenAI-compatible, so an existing voice app swaps an endpoint and a model name and keeps its event handlers. Mini, the conversational one built for assistants and tutors, isn't available under Zero Retention Mode or HIPAA, so the friendly model is the one a hospital can't put on its phone lines.

FLUX 3 Max is a video model from fal that renders about as fast as the clip plays: ten seconds of 720p in roughly twelve, against a little over two minutes for the FLUX 3 base it was post-trained from, at four cents a second until November 7. It's tied for first on Artificial Analysis's video arena, where Seedance 2.5 holds the same score and took four and a half minutes in fal's own tests. Twelve seconds is short enough to watch the motion, decide it's wrong and run it again, where the wait used to mean writing one careful prompt and going off to do something else while it rendered.

Plango is a Mac security app for people running coding agents, and the best thing in it is the oldest feature on the Mac: give a file a Finder color tag and Claude Code can't open it. Red means no. It'll also move .env secrets into Keychain, watch where agents send traffic and block a destination before they reach it, and scan whatever skills and plugins are sitting in your agent directories for things that shouldn't be there. Fred Rohrer and Rich built it around the lethal trifecta, Simon Willison's name for an agent holding your private data, a way to talk to the outside and something untrusted in the middle, and the model making those calls runs on the device rather than phoning anything home.