Google's deep-research-pro-preview-12-2025 agent has shut down: here's how to migrate
The deep-research-pro-preview-12-2025 managed agent shut down on 23 October 2026. API calls now fail. Here is how to migrate.
If you have any application making calls to deep-research-pro-preview-12-2025 through the Gemini API, those calls are now failing. Google’s hard shutdown date for that managed agent passed on 23 October 2026, and the endpoint is gone. There is nothing gradual about this, the model is off, and requests to it return errors.
Here is what you need to know to get back up and running.
What happened
Google deprecated deep-research-pro-preview-12-2025 on 8 October 2026, giving developers fifteen days before the hard shutdown on 23 October 2026. The agent was the Gemini API’s first managed Deep Research agent, launched in December 2025. It was designed to autonomously plan, execute, and synthesise results for multi-step research tasks, and it served that purpose through most of 2026.
Fifteen days is a tight turnaround for a production migration. Google’s deprecations page notes that the dates listed are the earliest possible shutdown dates and that Google communicates exact shutdown dates in advance, but fifteen days is still a short window for teams running research pipelines in production.
What you need to change
The fix itself is straightforward. In your interactions.create requests, update the agent parameter from:
deep-research-pro-preview-12-2025to one of two replacement agents:
deep-research-preview-04-2026, optimised for speed, designed to be streamed back to a client UI in real time.deep-research-max-preview-04-2026, optimised for comprehensiveness, suited to automated pipelines where thoroughness matters more than latency.
That is the entire code change. One parameter value, updated wherever you reference the old agent string.
Which replacement agent should you pick?
The choice depends on how you are using Deep Research in your application.
If you are building something user-facing where a person is waiting for results, deep-research-preview-04-2026 is the right choice. It is built to stream incrementally, so your UI can show progress rather than a blank screen while the agent works.
If you are running automated research pipelines, nightly data gathering, or any workflow where a human is not watching a spinner, deep-research-max-preview-04-2026 gives you the more thorough output. The tradeoff is that it takes longer, which matters less when nothing is waiting on a response in real time.
Neither replacement agent has a shutdown date announced on the deprecations page as of the time of writing, which gives you more runway than the original agent offered.
Where to look in your codebase
Search your codebase for the string deep-research-pro-preview-12-2025. It will appear wherever you construct an interactions.create request. That is the only place the change needs to happen. The rest of the Interactions API surface, the authentication, the request structure, and the response handling, remains the same.
The Interactions API has been Google’s default interface for Gemini models and agents since June 2026, so if your integration pre-dates that, it is worth checking whether you are already on the interactions.create endpoint or still using an older path.
A note on the broader deprecation wave
The deep-research-pro-preview-12-2025 shutdown was part of a wider set of retirements. Three Veo 3.1 preview models and gemini-omni-flash-preview shut down on 22 October 2026, one day earlier, and five speech and live audio previews are scheduled to follow on 17 November 2026. The same 8 October 2026 changelog entry that deprecated the Deep Research agent also deprecated two Gemini Flash models: gemini-3.7-flash replaced by gemini-3.8-flash, and gemini-3.5-flash replaced by gemini-3.6-flash.
If your application touches any of those models, it is worth checking Google’s deprecations page and changelog now rather than waiting for the next failure.
What this means for you
If your application is broken right now because of this shutdown, the path forward is clear: update the agent parameter and redeploy. The two replacement agents are well-suited to the same use cases the original covered, and the split between speed and comprehensiveness gives you a useful handle for optimising your specific workflow.
If you are not affected yet but you are running any Gemini API integration, it is worth bookmarking the deprecations page. Google has been moving through preview model retirements at a steady pace, and preview agents in particular tend to have shorter lifecycles than stable GA models. Keeping an eye on the page means you will not be caught out by a fifteen-day window again.