Gemini API Deep Research preview shuts down today: check IDs

If you wired Google's research agent into a workflow last winter, check it today. The Gemini API Deep Research preview agent, deep-research-pro-preview-12-2025, has an October 23, 2026 shutdown date. Yesterday, four Veo 3.1 and Omni preview models hit theirs. On November 17, five more preview audio and voice models go. A shut-down preview endpoint does not degrade politely. It stops answering.

What actually happened

Google's Gemini API changelog for October 8 announced the deprecation: the December 2025 Deep Research agent shuts down October 23, and the replacements are deep-research-preview-04-2026 and deep-research-max-preview-04-2026, both released April 21.

The deprecations page lists the rest of the queue:

• October 22: veo-3.1-generate-preview, veo-3.1-fast-generate-preview, veo-3.1-lite-generate-preview and gemini-omni-flash-preview
• October 23: deep-research-pro-preview-12-2025
• November 17: gemini-3.1-flash-tts-preview, gemini-3.1-flash-live-preview, gemini-2.5-flash-native-audio-preview-12-2025, gemini-2.5-flash-preview-tts and gemini-2.5-pro-preview-tts

Google calls these dates the "earliest possible" retirement dates, so a model can live a little longer. Do not plan on it.

The migration is mostly a string change. Per the Deep Research docs, both new agents run through the Interactions API (client.interactions.create) with background=True, not generate_content. The standard agent is built for speed and streaming to a UI. Max is built for depth. The new agents also take MCP servers with an allowed_tools list, File Search stores, and a collaborative_planning flag that returns a plan for you to approve before the agent runs the full job.

Why Gemini API deprecations matter for your business

We see the same failure every quarter. Someone builds a useful automation on a preview model ID, it works for months, and nobody owns the date. Then a nightly report or a product-video job fails with a 404, and the first person to notice is a customer.

Grep for model IDs today. Search your code, your env files and your no-code tools for preview in any model string. Every hit needs an owner and a date.

Put model IDs in config, not code. One file, one place to change, one deploy. If a shutdown means editing six services, your architecture has a cost you will pay again.

Test the replacement, do not just swap it. Deep Research Max will not cost or behave like the December agent. Run five real research jobs on both new agents, compare output and run time, then choose. Turn on collaborative planning for anything a client will read.

Add an alert on model errors. A 404 from a model endpoint should page someone, not sit in a log.

Sources: Google Gemini API changelog, Gemini API deprecations, Gemini Deep Research docs.