Google launched “Gemini Deep Research” its deepest AI research agent yet — on the same day OpenAI dropped GPT-5.2

• Reimagined research agent built on Gemini 3 Pro (December 2025)
  • Google announced the updated Gemini Deep Research on Thursday, rebuilding it on Gemini 3 Pro, described by Google as its “most factual” foundation model.
  • The focus is reliability in long-running, multi-step reasoning tasks, where hallucinations can invalidate entire outputs.
  • Gemini 3 Pro is optimized to minimize factual errors during extended agentic workflows.
  • The developer-facing Deep Research Agent is explicitly described as an agent that autonomously plans, executes, and synthesizes multi-step research and produces detailed, cited reports.
  • Research runs are designed to take minutes, not seconds, reflecting an “analyst-style” workflow rather than low-latency chat.
  • In product messaging around Deep Research, Google frames the feature as an early demonstration of “agentic” capabilities becoming practical in everyday workflows, not just a model demo.
• Developer embedding via the new Interactions API
  • Google introduced the Interactions API, allowing developers to embed Gemini Deep Research capabilities directly into their own applications.
  • This moves the product beyond static research reports toward customizable, agent-driven workflows.
  • The API reflects Google’s strategy for the “agentic AI era,” where AI systems act autonomously across tools and services.
  • The Gemini Deep Research Agent is in preview and is exclusively available via the Interactions API; it cannot be accessed through generate_content.
  • Long-running execution is a first-class requirement: developers must run the agent asynchronously using background=True, then poll the interaction until the state transitions from in_progress to completed (or failed).
  • The consumer Deep Research experience is described as operating “under your supervision,” beginning with a multi-step plan that the user can revise or approve before the system executes.
• Large-context synthesis for professional research use cases
  • Gemini Deep Research is designed to process very large context inputs and synthesize extensive information collections.
  • Google reports customer usage in areas such as corporate due diligence and drug toxicity safety research.
  • These tasks require sustained reasoning across many documents and decision points.
  • The Deep Research Agent uses iterative cycles of planning, searching, reading, and writing; this multi-step loop typically exceeds standard synchronous API timeout limits.
  • Deep Research can work with both public web information and user-provided data:
    • By default, it can use public internet information through tools including google_search and url_context.
    • Access to a user’s own data can be added via File Search (file_search), which is described as experimental for Deep Research usage.
  • Google positions the consumer version as a way to compress “hours” of manual tab-juggling into “minutes,” by repeatedly searching, learning, launching follow-on queries, and synthesizing the findings into a single report with links to sources.
  • The consumer Deep Research write-up also illustrates typical user scenarios beyond enterprise research, including:
    • A graduate-student-style research sprint (example: preparing a robotics presentation on autonomous vehicle sensor trends).
    • Small business planning (competitor analysis and potential location recommendations).
    • Marketing planning (researching recent AI-powered campaigns to benchmark for 2025 planning).
  • The earlier rollout connects Deep Research performance to a 1M token context window, paired with “advanced reasoning capabilities,” as a foundation for long, readable research reports.
• Planned integration across Google products
  • Google confirmed upcoming integrations with Google Search, Google Finance, the Gemini App, and NotebookLM.
  • The company frames this as preparation for a future where AI agents conduct information retrieval instead of users manually searching.
  • This positions Deep Research as infrastructure, not just a standalone tool.
  • In the earlier product rollout, Google specified rollout sequencing: Deep Research launched in Gemini Advanced (initially English) on desktop and mobile web, with availability planned for the mobile app and Workspace accounts in early 2025.
  • The same rollout emphasized practical workflow integration by allowing users to export Deep Research reports directly into a Google Doc, preserving citations/links for verification and follow-up.
  • Important product-to-API difference: the consumer experience described a user-supervised workflow where the system proposes a multi-step plan for revision/approval, while the developer Deep Research Agent lists human-approved planning and certain structured-output controls as current limitations.
• New open-source benchmark: DeepSearchQA
  • Google introduced DeepSearchQA, a new benchmark designed to test complex, multi-step information-seeking agent tasks.
  • The benchmark is open sourced, allowing external evaluation and comparison.
  • It complements existing benchmarks focused on agent reasoning rather than single-turn accuracy.
  • Google also tested Deep Research against Humanity’s Last Exam (an independent general-knowledge benchmark described as containing highly niche tasks) and BrowserComp (a benchmark for browser-based agentic tasks).
• Benchmark performance versus competitors
  • Google reports that Gemini Deep Research leads on DeepSearchQA and the independent Humanity’s Last Exam benchmark.
  • OpenAI ChatGPT 5 Pro placed a close second on both.
  • OpenAI slightly outperformed Google on BrowserComp, a benchmark for browser-based agent tasks.
  • The comparisons were presented as a snapshot and immediately pressured by fast-moving releases across the frontier-model ecosystem.
• Competitive timing with OpenAI GPT-5.2 (Garlic)
  • Google’s announcement coincided with OpenAI’s release of GPT-5.2, codenamed Garlic.
  • OpenAI claims GPT-5.2 outperforms rivals across multiple benchmarks, including OpenAI’s internal tests.
  • The near-simultaneous releases highlight intensifying competition in high-end agentic AI systems.
  • Google’s broader “agentic” positioning has been consistent across product messaging: Deep Research was described as an early, concrete step toward assistants that can act on a user’s behalf, not just answer questions.
• Operational mechanics developers must account for
  • Background execution: Developers must set background=True for Deep Research tasks and poll using the returned interaction id. Agent execution in background also requires store=True.
  • Streaming updates: Real-time progress updates are supported when stream=True and background=True. For intermediate progress and step-level updates, developers must enable thinking summaries in agent_config (if not set to "auto", streaming may only yield final output).
  • Resilience: Streaming implementations are expected to handle network interruptions by resuming with two values: the interaction ID and the last processed event ID.
  • Follow-up interaction: Developers can ask clarifying questions or request elaboration on specific report sections using previous_interaction_id, continuing the thread without re-running the entire research job.
  • Steerability: Output can be shaped through explicit formatting instructions in the prompt (sections/subsections, comparative tables, or tone such as “technical” vs “executive”), though fully structured output support is also listed as a current limitation in the agent documentation—making prompt-based formatting guidance the practical control surface today.
  • Time bounds: The agent has a maximum research time of 60 minutes, with most tasks expected to complete within 20 minutes.
  • In the consumer experience description, the research process is characterized as completing “over the course of a few minutes,” with the system repeatedly running search-and-read loops and then producing an organized report with direct links to sources.
• Availability, pricing, and safety constraints
  • The Deep Research Agent is available through the Interactions API in Google AI Studio and the Gemini API.
  • A time-bound pricing note applies to tool usage: Google Search tool calls are free until January 5, 2026; standard pricing applies after that date.
  • Safety considerations emphasized for web + file access:
    • Uploaded files can contain prompt-injection attempts; only trusted documents should be provided.
    • Public web browsing can encounter malicious pages; citations should be reviewed to verify sources.
    • Sensitive internal data should be handled carefully, especially when web browsing is enabled.
  • Current limitations include: Interactions API public beta (schemas may change), no custom function-calling tools or remote MCP servers for the agent, no audio input support, and constraints/restrictions on grounded Google Search results.
  • The consumer product update pairs Deep Research with an “experimental model” track (Gemini 2.0 Flash Experimental) and explicitly notes that early-preview models may be incompatible with some Gemini features.

Why This Matters

As research agents become embedded across products, APIs, and workflows, the competitive question is no longer who has the smartest chatbot, but who controls the most reliable AI infrastructure. Google’s push toward long-running, asynchronous research agents reflects a future where AI operates quietly in the background, executing complex tasks at scale, reshaping how information work is done across enterprises, platforms, and everyday tools.

This article was drafted with the assistance of generative AI. All facts and details were reviewed and confirmed by an editor prior to publication.