Overview
- Google released the Interactions API in public beta via Google AI Studio, unifying access to Gemini models and built-in agents through a single /interactions endpoint.
- A reimagined Gemini Deep Research agent is now available to developers, using Gemini 3 Pro and training aimed at reducing hallucinations during complex, multi-step research.
- The API adds features for agentic apps, including optional server-side state, interpretable agent histories, background execution for long jobs, and remote Model Context Protocol tool support.
- Google reports state-of-the-art results for the agent at 46.4% on Humanity’s Last Exam, 66.1% on DeepSearchQA, and 59.2% on BrowseComp, alongside lower-cost report generation.
- Google open-sourced the DeepSearchQA benchmark with 900 causal-chain tasks across 17 fields and said Deep Research is coming to Search, NotebookLM, Google Finance and the Gemini app, with Vertex AI support planned, as third-party coverage noted the same-day debut of OpenAI’s GPT‑5.2.