Kimi K3.1 Pro launches, frontier models cut off live internet training

The Qwen app kicked off invite testing for 'My Apps.' With zero coding needed, users can describe their ideas in plain language to build custom tools that store data and update content, useful for organizing events or planning trips. Once connected with cloud permissions, it can also pull in data from platforms like Amap and Damai.

Grok Bot added a 1Password account integration. Once users approve and share the login credentials they want, the agent can log into third-party sites or services for them. The feature will roll out to users over the next few days.

Anthropic announced a unified experience for Claude Enterprise, letting users switch from everyday Q&A to cloud agent tasks in the same chat window. These tasks run in isolated cloud virtual machines and keep running even if you shut down your computer, with progress sync available on mobile.

Developer Simon Willison shared a prompt trick that has Claude Opus 5.5 High generate a native web canvas. When a user uploads a photo, the AI uses the Sobel edge-detection algorithm to trace outlines in seconds and redraws them as a sketch on the page with simulated brushes, with zero third-party plugins needed.

Moonshot AI launched Kimi K3.1 Pro early on Oct 24, featuring around 3 trillion parameters and a 1M context window, ranking fourth on the Artificial Analysis overall leaderboard. The team also released a Highspeed version that is 2 to 3 times faster, and committed to opening the full model weights by Nov 4.

Kuaishou officially released its next-generation video model, Kling 4.0, which generates clips up to 30 seconds long and supports up to 10 keyframe controls. The newly added Omni Reference feature lets you combine up to 15 image, video, subject, or audio references in a single task, and the model API is already live.

MiniMax announced it has open-sourced the weights for MiniMax M3.1 Flash. The model has 428B total parameters with 23B actively running. The company stated that capabilities across the board are improved over the previous-generation M3 model.

OpenAI announced it will no longer connect to the open internet when training future frontier models, shifting completely to indexed caches and simulated data. The decision follows a September incident where a model accessed the web on its own through a configuration flaw. Anthropic set similar restrictions two weeks ago, as top labs grow wary of safety risks from models running unchecked on the open web.

Google updated its privacy policy, adding an exception to its past pledge not to use personal data for advertising: if users connect Gmail, Google Drive, or Photos to AI Search and other tools via Personal Intelligence, that content may now be used to deliver targeted ads.

The General Office of the State Council issued the 'Opinions on Adapting to the Development of Artificial Intelligence Technology to Help Build an Employment-Friendly Development Pattern.' It explicitly calls for cushioning the impact of AI job displacement, directing the technology primarily toward labor shortages and hazardous environments, carefully pacing its rollout in labor-intensive sectors, and encouraging enterprises to establish unemployment buffer periods and transition support.