Airbyte Agents cuts API round trips by keeping a live index of your business data. Instead of querying multiple APIs at runtime, agents run searches under 500 ms against that index, slash tool calls by 40%, reduce token use 80%, and drop multi-source query costs by 90%. Mistral OCR 4 packs state-of-the-art document intelligence into a single container. It handles 170 languages, spits out structured text with bounding boxes and confidence scores, and beats rivals by four times on speed—especially on low-resource scripts. Anthropic’s new Claude Tag lives in Slack channels, lets teams assign tasks, hook into codebases and external tools, and keeps thread-to-thread context. Their product team uses it for code generation, data analysis, support tickets, even debugging. ByteDance shipped Seedance 2.5, an AI video model that turns a prompt plus up to 50 reference images, clips or audio files into a 30-second 4K video. More references mean finer control. It’s coming to China next month; global availability isn’t set yet. OpenAI quietly rolled out Bidirectional Voice Mode on ChatGPT, powered by its Bidi 1 audio model. It can speak, listen and hold a back-and-forth, pivot mid-conversation—and even sing or beatbox under strict copyright limits. Security and orchestration got deep dives. Gray Swan’s team mapped out indirect prompt injections—jailbreaks that exploit how LLMs can’t tell system tags from user input. IBM’s CUGA framework standardizes agentic app building with built-in state management, configurable reasoning modes and policy controls. NVIDIA’s Agent Toolkit helps enterprises spin up specialized agents over open models, tools and secure runtimes—companies like CrowdStrike and Cadence are already on board. Engram keeps AI models “warm” on your private context—documents, chats and code—so they don’t waste cycles rereading the same files each session. On the open-source front, Graphsignal profiles inference at scale with minimal performance hit; Krea 2 expands text-and-image creative control via a prompt expander; Unlimited OCR tackles long‐form document parsing with a constant KV cache.