Enterprise AI Agents, Trillion-Parameter Models and Decision APIs

Google Cloud introduced the Gemini Agent at Gemini at Work 2026. The agent operates through a single prompt interface to conduct research, coding, and writing across enterprise platforms including Slack, Salesforce, Jira, and Google Workspace. Tasks run in the cloud with persistent execution, continuing after users disconnect. It supports multi-model orchestration across Gemini and Anthropic Claude models, with security controls like data encryption and human approval gates for critical actions.

Mistral AI introduced Mistral Large 4, a multimodal mixture-of-experts model with 1.05 trillion total parameters and 49 to 52 billion active parameters per token. The model includes a 1 million token context window and scored 61.7 percent on DeepSWE v1.1 alongside 49.8 percent on the Coding Agent Index. Mistral trained the system on approximately 3,800 Nvidia chips and scheduled open weights for release at the end of October 2026.

OpenAI launched the Decisions API running on the GPT-6 Luna model for fast evaluation tasks without web search. The API operates roughly ten times faster than the standard Responses API, achieving response latency of approximately 270 milliseconds. It handles three query formats: true or false predicates, classification choices, and ordered scoring levels, with input pricing around 0.03 dollars per 1,000 images.

Anthropic released a Google Workspace add-on that embeds Claude inside Google Docs, Sheets, and Slides. Users can initiate tasks from a sidebar in Google files or paste file links inside Claude to edit documents and spreadsheets directly. The integration includes Ask Before Edits with approval cards alongside an Accept All Edits mode, and requires paid team or enterprise account configurations.

Google released the Stitch CLI and Stitch MCP to bridge AI design tools with terminal-based coding agents like Claude Code, Cursor, and Antigravity. Installed through npm, the CLI lets developers generate screens, capture live screenshots, and automate UI generation without a web interface. The system uses DESIGN.md files to store design tokens and maintain consistent frontend styling across project codebases.

Google launched EmbeddingGemma 2, a lightweight 740-million parameter multimodal embedding model that processes text, images, video, and audio in a unified vector space. The architecture uses an 8,000 token context window and runs on-device, consuming between 191 megabytes for text up to 567 megabytes for multimodal tasks. Matryoshka Representation Learning allows compressing vector dimensions from 768 down to 128, cutting index storage requirements up to sixfold.