Gemini Deep Research's old agent retires October 23: change agent

Which entry point shuts down on October 23?

Google's October 8 Gemini API release notes say deep-research-pro-preview-12-2025 is deprecated and will shut down on October 23. At the October 9 check, this remained a future plan. The announcement gives no exact time or time zone.

The deprecation table also lists October 23, but explains that its dates generally indicate the earliest possible retirement date. Plan this migration using the explicit release-note notice as well. The scope is a managed research agent in the Gemini API, not the consumer Gemini app, every Gemini model, or the entire research feature.

Choose speed or more comprehensive research

The standard replacement is deep-research-preview-04-2026. Google positions it for speed and efficiency, particularly when research progress is streamed to a client interface. If the main goal is a timely research result for a user, start by evaluating this option.

The other option, deep-research-max-preview-04-2026, emphasizes more comprehensive information gathering and synthesis for automated research. Max is not a guarantee of better suitability for every task. Compare the required coverage, time and actual cost.

The deprecation table lists April 21 as the release date for both versions, with no announced shutdown date. The October 8 news concerns the old agent's retirement and migration, not the replacements' first release. The current research agents remain in Preview.

Change agent, not the model name

Find the interactions.create request that actually creates the research task, and replace the old identifier in agent with your chosen replacement. Deep Research searches, reads and synthesizes information as an agent. Google provides it exclusively through the Interactions API; do not pass it to generateContent as an ordinary model.

Also check fixed identifiers in scheduled jobs, wrapper functions, environment configuration and shared templates. For example, a webpage might use the replacement while a nightly summary script still keeps the old name; that script would need a separate change. This is an unexecuted example, not a fault found in your project.

Research is still running: do not create another copy

Google requires asynchronous execution with background=True, which also requires store=True. A task can perform multiple rounds of searching and reading rather than return a final report immediately like ordinary chat. Confirm your project's requirements for storing research interactions and input material before choosing what to submit.

After creation succeeds, save the returned interaction ID and use that same ID to poll the status or receive streamed updates. Read the report after completed and record the error after failed. If it is still running, keep tracking the original task; closing a webpage or reaching a waiting timeout does not automatically mean it failed.

If a streaming connection drops, Google's example reconnects using the original interaction ID and the last received event ID. Check the original task when its outcome is unknown, avoiding duplicate research and costs. If you have not received a task ID, you still cannot assume the service did not create it.

How should migration costs be checked?

Deep Research is billed according to the underlying Gemini models and the tools actually used. One research request can involve multiple rounds of planning, searching, reading and reasoning. Do not estimate its cost from a single ordinary-chat output, or treat 'more efficient' as a fixed price or spending cap.

Teams already authorized to make calls can compare coverage, citation quality, total time and billing using the same nonsensitive research question. Check account availability, regions, quotas and current rates before expanding use. This article did not call the API, execute a migration or verify any account's costs.

What order should the migration follow before the deadline?

First inventory workflows that still use the old identifier, recording entry points, configuration locations, owners and IDs of existing in-flight tasks. Then choose a replacement based on delivery speed or research coverage. Keep versions of the original prompts and configuration rather than changing all tools and task requirements at once.

Next check the active agent configuration, background status polling and result retrieval. When the report arrives, verify that its citations support its conclusions and distinguish confirmed facts, forecasts and missing information. A completed status means the task ended; it does not guarantee accurate sources or a report suitable for direct publication.

Finally confirm costs and error handling, then gradually resume automated tasks. Pause expansion and retain records if key information is missing, citations are misleading or costs exceed the team's budget. A fallback after October 23 cannot depend on the old agent scheduled for shutdown.