OpenAI opened the Decisions API in public beta. Backed by GPT-6 Luna, it answers yes/no, multiple choice and score questions over text or images, and OpenAI claims up to 10x faster responses than the Responses API. Read: OpenAI opened the Decisions API in public beta. Backed by GPT-6 Luna, it answers yes/no, multiple choice and score questions over text or images, and OpenAI claims up to 10x faster responses than the Responses API. Read: Mistral opened API preview access to Large 4, a multimodal mixture-of-experts model with 1T total and 49B active parameters. Open weights are due at the end of October; Artificial Analysis scores it 38 on its Intelligence Index. Read: GitHub says agent-driven development is straining Spokes, where every replica joins each write. It reported 7.38 billion commits in September and plans to separate durability from read scale. Read: EmbeddingGemma 2 is a 740M-parameter Apache 2.0 embedding model on the Gemma 4 architecture. It maps text, code, images, audio and video into one shared space for on-device search and RAG. Try: A placebo-controlled test ran nine top-starred Claude Code skills on SWE-bench Verified, Terminal-Bench and TBLite. Only two beat neutral text of the same length, one did worse, and none lowered cost. Read: Anthropic merged Project Glasswing and its Cyber Verification Program into three tiers for verified security professionals. Defense and Red Team tiers unlock authorized pentesting on its top models; physical-harm actions stay blocked. Read: Claude for Google Workspace is in beta on all paid plans. A sidebar in Docs, Sheets and Slides reads the open file and proposes edits for approval, using the user's connectors and skills.
The Safety Check That Runs on the Conversation, Not the Turn