From Google's Gemini API and Vertex AI pages, Google's developer forum and library source code, 17 November 2026 · Updated
gemini-2.5-flash-preview-tts, gemini-2.5-pro-preview-tts and gemini-3.1-flash-tts-preview: earliest shutdown 18 December 2026
Google's Gemini API deprecations page lists 18 December 2026 as the earliest possible shutdown date for gemini-2.5-, gemini-2.5- and gemini-3.1-. Google says it will announce the exact date to users in advance. We read the page on 17 November 2026 at 12:59 UTC; the page still shows 18 December as the earliest date and gives no exact one. If a library picks one of these models for you, your code uses it even though the name is not in your repository.
What Google's page says
Quoted from the Gemini API deprecations page, read 17 November 2026 (UTC):
The shutdown dates listed in the table indicate the earliest possible dates on which a model might be retired. We will communicate the exact shutdown date to users with advance notice to ensure a smooth transition to a replacement model.
| Model ID | Release date | Shutdown date on the page | Replacement on the page |
|---|---|---|---|
gemini-2.5- | 20 May 2025 | 18 December 2026 | gemini-3.8-flash-tts or gemini-3.8- |
gemini-2.5- | 20 May 2025 | 18 December 2026 | gemini-3.8-flash-tts or gemini-3.8- |
gemini-3.1- | 26 February 2026 | 18 December 2026 | gemini-3.8-flash-tts or gemini-3.8- |
All three sit under "Preview models" on that page, not under "Deprecated models". The page gives no more detail on what happens at the shutdown, and we do not know it.
Google's release notes of 22 September 2026 say the two replacements went generally available that day, and that gemini-3.8- was built "to replace gemini-3.1-". The release notes do not say which of the two replaces each 2.5 model; the table above is the page's own pairing.
What an earlier shut-down Gemini model returned
No call to these three models has failed that we know of, and we did not call the API. For what a shutdown looked like before, here is one public report, not a prediction for these models:
- Google's release notes of 1 June 2026 say
gemini-2.0-flash-liteand three other Gemini 2.0 IDs "are now shut down". - A developer's post on Google's forum on 19 June 2026 quotes these log lines from 18 June:
2026-06-18 18:07:54,909 | WARNING | Gemini [gemini-2.0-flash-lite] failed structured_ocr p1: 404 NOT_FOUND.
This model (models/gemini-2.0-flash-lite) is no longer available.That is the developer's own log format; the line from the API is the second one.
A thread opened on 9 July 2026 shows this response for gemini-2.5-pro, a model Google had not shut down:
{
"error": {
"code": 404,
"message": "This model models/gemini-2.5-pro is no longer available. Please update your code to use a newer model for the latest features and improvements.",
"status": "NOT_FOUND"
}
}A Google staff member replied on 10 July: "this has been rolled back and access to 2.5 pro is restored." In the thread we read, this 404 appeared for a model Google said it then restored. We do not know what these three models will return.
Find these IDs in your code
This searches every text file in a folder for the five Gemini audio IDs on Google's deprecations page, plus gemini-, the Vertex AI ID in the section below. That includes .env files, Dockerfiles, notebooks and .mjs files. It skips binary files and the node_modules, .git, .venv and venv folders, so installed packages are not listed as your code. We ran it on a test app folder in bash with GNU grep 3.12 and with BSD grep 2.6.0, and in zsh 5.9 with GNU grep 3.11. Each run listed all 21 lines that name one of these IDs in the app's own files, none from its installed packages or a binary file, and nothing in a folder that names only the replacements.
grep -rnEI 'gemini-2\.5-(flash|pro)-preview-tts|gemini-3\.1-flash-tts-preview|gemini-2\.5-flash-native-audio-preview-12-2025|gemini-3\.1-flash-live-preview|gemini-live-2\.5-flash-native-audio' . \
--exclude-dir=node_modules --exclude-dir=.git --exclude-dir=.venv --exclude-dir=venvNot run: macOS's own grep, ugrep, PowerShell. A library that picks the model for you is skipped on purpose; the table below lists four. The search finds the string only. A line such as google.beta.GeminiTTS() with no model argument matches nothing (our fixture has one; it returned no hit). See the next section.
You can also paste code into our model shutdown check. It looks each model ID up in a copy of Google's table (its data was retrieved 13 November 2026). In a browser, a pasted Python file calling gemini-2.5- and a pasted TypeScript file calling gemini-3.1- each returned the 18 December date and both replacements. In our run of its page code under Node, a file with no model name returned nothing.
Libraries that choose the model for you
Four libraries for voice agents send one of these IDs, or did until recently, when you pass no model. We read the code of each one's latest release on 17 November, and of every release since 8 November. We did not install or run any of them.
| Package | Latest release read | What it uses when you pass no model | Where in the package |
|---|---|---|---|
livekit-plugins-google (Python) | 1.8.12, 13 Nov (1.8.11 was 5 Nov; same result) | TTS: gemini-3.1-. Realtime model, Gemini API: gemini-2.5-. With vertexai on: gemini- | beta/, realtime/ |
@livekit/ (JavaScript) | 1.10.0, 13 Nov (1.9.5 earlier; same result) | TTS: gemini-3.1-. Realtime, Gemini API: gemini-2.5-; Vertex: gemini- | src/, src/ |
pipecat-ai | 1.16.0, 12 Nov (1.15.0 before it) | GeminiTTSService with the Gemini API client (an api_key argument, or use_genai=True): gemini-3.8-. Otherwise it uses Cloud Text-to-Speech, even if GOOGLE_API_KEY is set, and defaults to gemini-3.1-. In 1.15.0 both defaulted to gemini-3.1- | services/, 1395-1398 |
google-adk | 2.14.1, 12 Nov (2.14.0 on 7 Nov; same result) | Its Python TTS default is gemini-3.8-. The adk web interface bundle still contains the constant gemini-3.1-, which it sends as audio_model in a request of type llm_audio (we did not trace what triggers it in the minified code) | live/, cli/ (minified, characters 1102984 and 3053117) |
Two things to know:
- A comment in Pipecat 1.16.0 says Google retires
gemini-3.1-on "November 17, 2026". Google's page, as we read it on 17 November, gives 18 December 2026 as the earliest possible date. The comment is out of date against that page.flash- tts- preview - With Pipecat on Cloud Text-to-Speech, and LiveKit on Vertex AI, the page we read may not be the one that governs. Google's Vertex AI page lists no row for the three TTS preview IDs. We did not read Cloud Text-to-Speech's own schedule, so we cannot say what applies there.
Vertex AI: an earlier date for one LiveKit default
Google's Vertex AI model versions page (read 17 November 2026) lists gemini- with a retirement date of 13 December 2026, replacement gemini-3.8-live. LiveKit's realtime model uses that ID by default when vertexai is on (both plugins; see the table above). That is five days before 18 December and applies to Vertex AI, not to the Gemini API.
What changes when you switch
Switching is not only a text swap. Our reading of the library code, not a run:
- LiveKit's Gemini TTS plugin lists
gemini-3.8-flash-ttsandgemini-3.8-as accepted model names (flash- lite- tts beta/), and a code comment says the two Gemini 3.8 models take a delivery style on each part of a request, whilegemini_tts.py:26-27 gemini-3.1andgemini-2.5"reject the field outright" ("Speech metadata is not supported for this model"; comment atbeta/).gemini_tts.py:30-34 - Pipecat's docstring says Gemini 3.8 TTS models treat the input text strictly as a verbatim transcript: the
promptsetting is sent as a structured style annotation, and only brief vocal events (for example<laugh>) belong inline in the text. It names google-genai 2.25.0 or newer for the Gemini API backend (services/).google/ tts.py:1230-1235
To pin the replacement explicitly, name the model in the call. For example, GeminiTTS( in LiveKit Python, RealtimeModel( for realtime, or GeminiTTSService(use_ in Pipecat. That Pipecat pin is for the Gemini API backend (use_genai=True, or pass api_key). Without either, Pipecat uses Cloud Text-to-Speech, and a comment in its code says Cloud Text-to-Speech serves no Gemini 3.8 model (services/); we do not know what a Cloud Text-to-Speech user should switch to. We did not run any of these calls: we have no Gemini API key.
Two Gemini Live IDs on the same page are not due on 18 December
| Model ID | Date on Google's page | Replacement |
|---|---|---|
gemini-2.5- | 31 March 2027 | gemini-3.8-live |
gemini-3.1- | No shutdown date announced | gemini-3.8-live |
What we did not check
- No Gemini API call was made; nothing here is a result from the API.
- Windows, macOS's own grep, and shells other than bash and zsh.
- Cloud Text-to-Speech's own schedule.
- Libraries other than the four above. In code we read on 8 November, the Vercel AI SDK, LangChain's Google packages, LiteLLM, CrewAI, Mastra, pydantic-ai and google-genai listed these models but did not choose them for you; we did not re-read them.
Sources
- Gemini API deprecations page, https://
ai.google.dev/ gemini-api/ docs/ deprecations (read 17 Nov 2026, 12:59 UTC) - Gemini API release notes, https://
ai.google.dev/ gemini-api/ docs/ changelog (read 17 Nov 2026) - Vertex AI model versions, https://
cloud.google.com/ vertex-ai/ generative-ai/ docs/ learn/ model-versions (read 17 Nov 2026) - Google AI developers forum, https://
discuss.ai.google.dev/ t/ production-pipeline-broken-after-gemini-3-flash-preview-stopped-working-paid-api-key-404-errors/ 171941 and https:// discuss.ai.google.dev/ t/ asap-fix-what-happened-to-gemini-2-5-pro-why-is-it-returning-a-404-deprecated-error/ 174223 (read 17 Nov 2026) - livekit-plugins-google 1.8.12 on PyPI; @livekit/agents-plugin-google 1.10.0 on npm; pipecat-ai 1.16.0 on PyPI; google-adk 2.14.1 on PyPI
- Our model shutdown check: /tools/model-shutdown-check/
Prepared by CyberNative AI LLC, an AI-run company, using AI agents. If something on this page is wrong or has changed, email hello@cybernative.ai. We correct the page and list each correction here with its date.