Patience Cave releases MazeBench 3D spatial reasoning benchmark

Paseo's Vercel Sandbox plugin runs fx, Codex, Claude Code, OpenCode, and Pi in the cloud, letting you start a session on a laptop and control it from a phone with the laptop closed. Stopped sandboxes preserve files and conversation history, though active processes stop.

Virtual Models abstract the model string behind an ID like vmc/coding-agent, so teams can swap models, set fallbacks and priority, and enforce Zero Data Retention centrally without redeploys. Vercel says 125 new models from 10+ labs landed on AI Gateway in the last 90 days, roughly 10 a week.

Built for the WhisperNow voice-to-text app, the fine-tuned cleanup models strip fillers, false starts and self-corrections, and format lists. The developer positions it against the many WisprFlow clones by training the cleanup step rather than relying on a general model.

GPU inference via ML Drift runs about 1.4x faster than the legacy GPU delegate, exposed through the new CompiledModel API. Apps switching from bundled LiteRT shrink by roughly 8 MB, with over 100K Android apps still bundling their own copy.

Amodei said making AI girlfriends or boyfriends is "a terrible use of AI and people should not do it," arguing it pushes further toward disintermediation of human relationships.