Gemini Deep Research Agent Now on Gemini API
Google launched a stronger version of its Gemini Deep Research agent for developers this week. It’s accessible right away through the new Interactions API using a standard Gemini API key from Google AI Studio. Their blog post calls it the first time devs can plug Google’s top research skills straight into apps.
Access and Setup
The agent lives in the Interactions API, Google’s unified endpoint for models and agents. Pick the model for basic stuff or set the “agent” parameter to “deep-research-pro-preview-12-2025” to fire up Deep Research. It’s in public beta, so expect tweaks, but it handles server-side state, background runs, and tool calls out of the box. Google’s API announcement lays out the docs and OpenAPI spec to get rolling fast.
9to5Google notes this dev preview drops today, with rollouts to Gemini app, Search, NotebookLM, and others coming soon.
Core Capabilities
This agent runs on Gemini 3 Pro, tuned to cut hallucinations and crank out solid reports on tough tasks. It plans queries, scans results, finds gaps, and loops back for more searches. Web navigation got a big upgrade to pull specifics from deep in sites.
- Takes your docs (PDFs, CSVs) plus web data for synthesis
- Lets you steer reports with prompts for structure, tables, headers
- Spits out detailed citations per claim
- Outputs JSON schemas for easy app parsing
Benchmark Scores
Google reports it hits 46.4% on the full Humanity’s Last Exam (beats Gemini 3 Pro’s 43.2%), 66.1% on their new DeepSearchQA benchmark (up from 56.6%), and 59.2% on BrowseComp (from 49.4%). They open-sourced DeepSearchQA too—a 900-task set across 17 fields testing multi-step web digs.
Real-World Use
Early tests show it aiding finance due diligence, biotech lit reviews, and market scans. Firms like Axiom Bio used it to speed drug toxicity predictions from papers. More features like charts and Model Context Protocol ties are on deck, plus Vertex AI support later.
