XNews

DeepMind/Google releases Gemini Deep Research & Interactions API

DeepMind/Google releases Gemini Deep Research & Interactions API

DeepMind launched Gemini Deep Research, a developer-focused agent built on Gemini 3 Pro and the new Interactions API. - Agent role: autonomous research planner and synthesizer. - Capabilities: plan, search, read, identify gaps, re-search. - Integration: combines private uploads with web information. - Model base: built on Gemini 3 Pro with multi-step RL training. - Verifiability: outputs include detailed citations/sources. - Benchmarks: DeepSearchQA 66.1%, Humanity's Last Exam 46.4%, BrowseComp 59.2%. - Interactions API: unified endpoint for models and agents, prompt-controllable outputs and JSON/table export. - Availability: developers can access via GoogleAIStudio and Gemini API keys. - Related: Disco/GenTabs uses Gemini 3 for generative tab-to-app experiences and richer browser workflows. - UX note: “Create image” option alters default tool-calling behavior in Gemini apps.

Key insights

Gemini Deep Research launched as an autonomous developer-grade research agent

DeepMind announced Gemini Deep Research as a research-oriented agent that autonomously plans search strategies, reads results, identifies information gaps, and performs targeted re-search. The agent combines uploaded private documents with public web data and is trained with multi-step reinforcement techniques atop Gemini 3 Pro. Outputs are designed to include detailed citations to support verifiability.

Interactions API provides a single interface to models and agents

DeepMind introduced the Interactions API as a unified developer endpoint to call both models and agents. The API supports prompt-defined report structures, JSON and table outputs for downstream processing, and tool integration. Deep Research is the first agent built on this API and is available through GoogleAIStudio with Gemini API keys.

Benchmarks show SOTA performance on multi-step research tasks

DeepMind published benchmark results for Gemini Deep Research: DeepSearchQA 66.1%, Humanity's Last Exam 46.4%, and BrowseComp 59.2%. These datasets target causal-chained research tasks, comprehensive information retrieval, and locating hard-to-find facts, with Deep Research reported as best-in-class on the published tests.

Gemini ecosystem expands into browser tooling and image tool semantics

Related developer tooling includes browser-integrated Disco/GenTabs that transforms collections of open tabs into generated interactive apps using Gemini 3 capabilities. Additionally, Gemini app UI semantics like a ‘Create image’ toggle change the model’s default tool-calling behavior, affecting whether image-generation tools are invoked by default.

Source posts