OpenAI's Python SDK 3.55 adds typed configuration for Live MCP servers, including tool allowlists and approval policies.
More tools are useful; knowing which ones can run is even better.
https://t.co/bHTca7LfLu
SenseNova's Loop-9B is out of preview: 32 unique layers execute 56 times by reusing weights. It takes text and images, with a 512K context.
Clever tradeoff: fewer weights still means more compute. Curious how it feels locally.
https://t.co/yEuo6kKJHC
https://t.co/tylQ0QZ0yh
OpenCode 1.19.6 now replaces the oldest images with placeholders when a session's image history exceeds 25 MiB.
Fewer oversized requests sounds good. I'd also want to know which screenshots the agent can no longer see.
https://t.co/Ue80DBLDyx
Qwen Code 0.28 defaults Web Shell sessions in trusted Git workspaces to separate worktrees, with branch creation and handoff.
Parallel agents sharing one checkout have caused enough drama already. Curious how handoff feels in practice.
https://t.co/Z38QuZhZ5B
Hermes GenUI 0.3 can turn MCP data into dashboards you pin in the sidebar.
A detail I like: refresh repeats the original tool calls instead of asking the model to reinvent the numbers.
https://t.co/xDpOoUYulG
Mistral's three-month free Vibe offer for Europe already ran out of codes for today. The company says to watch for more news tomorrow.
That was quick. Free AI deals really do move at internet speed.
https://t.co/1bMQJ2lvVg
HeyGen's Avatar 6 is now public. Its new model can follow written cues for gestures timed to specific words, and generate talking video from a single photo.
The thumbs-up on cue is fun. I'm curious how it handles a long, unscripted conversation.
Image: HeyGen's demo.
Luminary just introduced Autopilot for physics AI: agents can run data generation, training and inference from an engineering prompt.
I like the practical detail: it asks before changing data or spending compute, with approvals you can revoke.
https://t.co/7yTg2mCUpa
Microsoft says MAI Code 2 reaches 98% of GPT-6.1 Sol's performance at 40% of the cost across four coding benchmarks.
One detail: part of that comparison uses a 50% promotional price. Curious how it holds up after the promo.
Microsoft says it's working with 40+ organizations on AI transparency standards, including ways to independently verify agent activity.
That's a useful direction. An agent's own account of what happened shouldn't be the only audit trail.
https://t.co/45Rj394oqA
OpenSpec 2.1 can install coding-agent skills once for every project on your machine.
Less setup sounds great. One detail: hosted Copilot agents may still need the skills in the repo, so the upgrade lets you keep them project-local.
https://t.co/TqzxPd0lxK
GPT-6.2 Astra and Sol are out. OpenAI says Astra's API input and output prices are 20% lower than GPT-6 Astra's; Sol keeps 6.1 Sol pricing.
Cheaper per token is nice. I'm curious how much the cost per finished task changes in real use.
https://t.co/byLoKVxeP6
Open WebUI 0.13.4 brings local MCP servers into the chat through Tether, without opening a port.
The catch: those local commands run with your account's access, outside the workspace sandbox. Convenient, but definitely worth checking what you connect.
https://t.co/pup7lNPWzJ
GitHub MCP Server 2.2.0 adds opt-in Skills: 18 bundled workflows, filtered to the tools your connection actually has.
Nice idea. One detail to watch: enabling Skills replaces server instructions, even if your client doesn't support Skills.
https://t.co/N2Jv3mb5B4