Google Deep Research APIs

Google has turned its Deep Research tool into an application programming interface, letting developers call autonomous research agents from the Gemini API instead of only using them inside Google’s consumer apps. (blog.google) A research agent is software that plans a task, searches for sources, reads them, and writes a report; Google says its Gemini Deep Research Agent does those steps with web search and a user’s own data, then returns cited results after background processing that can take several minutes. (ai.google.dev) Google said on April 21, 2026 that it is adding two versions, Deep Research and Deep Research Max, and that Deep Research Max is available in public preview on paid Gemini API tiers. The company said the new versions build on a developer release from December 2025 that exposed the earlier Deep Research agent through the Interactions API. (blog.google 1) (blog.google 2) The Interactions API is Google’s newer interface for long-running agent workflows, replacing one-shot prompts with server-side state, background execution, tool calling, and support for the Model Context Protocol, a standard for connecting models to outside tools and data sources. Google says developers can use the same endpoint for raw Gemini models or specialized agents such as Deep Research. (developers.googleblog.com) (ai.google.dev) That puts a product Google first shipped to Gemini Advanced subscribers in December 2024 on a more direct path into software products. In that earlier rollout, Deep Research was framed as a consumer research assistant inside Gemini; the newer API framing treats the same kind of work as infrastructure developers can embed into apps and internal tools. (blog.google 1) (blog.google 2) Google’s developer docs say the current Deep Research agent is powered by Gemini 3.1 Pro, while the Interactions API docs list specialized agent IDs and note that the interface remains in beta, with features and schemas that may still change. That means developers can start building against the system now, but they should expect some churn while Google finalizes the platform. (ai.google.dev 1) (ai.google.dev 2) Google has been laying the groundwork for this shift for months. In December 2025, it introduced the Interactions API as a single endpoint for models and agents, and paired it with agent-building tools such as the Agent Development Kit and support for open-source frameworks including LangChain, LlamaIndex, Pydantic AI, n8n, and Vercel’s AI SDK. (blog.google) (developers.googleblog.com) The practical pitch is straightforward: instead of writing custom logic for search, memory, tool use, and report generation, a developer can hand a research job to Google’s managed agent and poll for the answer when it finishes. Deep Research Max is Google’s signal that it wants a premium tier for heavier-duty autonomous work, not just chat completions. (ai.google.dev) (blog.google)