Gemini API deadlines: Deep Research on 23 Oct, silent Flash re-routing, and the Gemini 2.5 split

Google’s Gemini API changelog has a habit of packing several important changes into one short entry. The 8 October update is a good example: two Flash models were deprecated and quietly redirected to newer ones, and the original Deep Research preview agent got a shutdown date of 23 October. Add the dates on Google’s deprecations page and there are three deadlines in the next five weeks.

Our retirements checklist from yesterday covered OpenAI, Anthropic and the Gemini app. This post is the developer side: the Gemini API, and Gemini on Google Cloud’s Agent Platform (formerly Vertex AI).

The calendar

From Google’s deprecations page and changelog:

Date	What stops working	Google’s suggested replacement
22 Oct	`veo-3.1-generate-preview`, `veo-3.1-fast-generate-preview`, `veo-3.1-lite-generate-preview`, `gemini-omni-flash-preview`	`gemini-omni-1.1-flash`
23 Oct	`deep-research-pro-preview-12-2025` (Deep Research agent)	`deep-research-preview-04-2026` or `deep-research-max-preview-04-2026`
17 Nov	`gemini-3.1-flash-tts-preview`, `gemini-2.5-flash-preview-tts`, `gemini-2.5-pro-preview-tts`	`gemini-3.8-flash-tts` or `gemini-3.8-flash-lite-tts`
17 Nov	`gemini-3.1-flash-live-preview`, `gemini-2.5-flash-native-audio-preview-12-2025`	`gemini-3.8-live`

One caveat: the deprecations page describes its dates as the earliest a model can be shut down, with the exact date announced with notice. Treat them as the day to be done by, not a day you might get extra time.

Deep Research: pick speed or depth

If you call the Deep Research agent, 23 October is the firmest date in this list. The changelog says deep-research-pro-preview-12-2025 “will be shut down on October 23, 2026” and offers two replacements:

deep-research-preview-04-2026, aimed at speed, suited to streaming results into a user interface.
deep-research-max-preview-04-2026, aimed at the most thorough research.

The choice isn’t only a rename. If your product shows Deep Research output to users while it runs, the faster agent is the closer fit. If it runs in the background and a report lands in someone’s inbox later, try the max version. Either way, rerun a handful of your real research prompts on both before the 22nd and compare length, time to finish and source quality.

The silent change: Flash auto-routing

The other two items in the 8 October entry won’t break anything, which is exactly why they’re easy to miss:

gemini-3.7-flash is deprecated and replaced by gemini-3.8-flash. Requests to the old name are automatically routed to the new model.
gemini-3.5-flash is deprecated and replaced by gemini-3.6-flash, also with automatic routing. Google also suggests gemini-3.8-flash as an option.

The deprecations page is explicit that these endpoints are not shutting down: they keep answering, but a different model is behind them.

For a chatbot, that may be fine. For a pipeline, it is a model change you didn’t ship. Things that can move when the model underneath changes:

Output format. Prompts tuned to produce a specific JSON shape or label set can drift.
Eval baselines. Scores you recorded last month were for a different model.
Latency and token use. Worth comparing against your own logs from before 8 October.
Cost. The changelog doesn’t say how routed requests are billed. Check your billing report rather than assuming.

Our suggestion: change the model string yourself, to the replacement Google names, and rerun your evals. Pinning the name you are actually getting makes the next change visible in your diff, not in production.

Gemini 2.5: two different answers depending on where you call it

This is where teams can get caught out, because Google’s own pages describe Gemini 2.5 differently for the two platforms.

On the Gemini API (Google AI Studio keys). The changelog entry of 18 September says Gemini 2.5 models are not deprecated and remain available until further notice, but only to users who have used them before. The deprecations page lists no shutdown date for gemini-2.5-pro, gemini-2.5-flash or gemini-2.5-flash-lite. For new projects, Google recommends 3.5 Flash-Lite or 3.8 Flash.

On Agent Platform (formerly Vertex AI). Google’s Firebase AI Logic model page says Gemini 2.5 Pro, Flash and Flash-Lite on the Agent Platform Gemini API “will shut down on October 20, 2026”, and that stable Gemini Live API 2.5 models are not affected. We couldn’t load the Agent Platform lifecycle table itself to confirm the date, so if you run 2.5 on Google Cloud, check the model page in your own console today.

In practice: if your production traffic goes through Google Cloud and still names a Gemini 2.5 model, assume you have about a week, not “until further notice”.

A one-hour audit

Search your code and config for deep-research-pro-preview-12-2025, gemini-3.5-flash, gemini-3.7-flash, gemini-2.5-, veo-3.1- and the TTS and Live preview names in the table above. Include environment variables and any model names stored in a database.
Note which platform each call goes through: Gemini API key, or Google Cloud project. The 2.5 answer depends on it.
Work by date: Gemini 2.5 on Agent Platform first (20 October), then the Veo and Omni previews (22 October), then Deep Research (23 October).
Pin auto-routed Flash calls to the model they are really using, and rerun evals.
Schedule the voice work for November: the TTS and Live previews have until 17 November, and their replacements are already generally available.

None of this is hard. The risk is in the names nobody remembers hard-coding, so start with the search.