Vertex AI Gemini retirements in November and December 2026
The dataset has 2 Google Vertex Gemini ids with an exact shutdown date in this window: gemini-3.6-flash, gemini-live-2.5-flash-native-audio. The model versions page lists gemini-3.6-flash with retirement date November 19, 2026 and replacement model ID gemini-3.8-flash. It lists gemini-live-2.5-flash-native-audio with retirement date December 13, 2026 and replacement model ID gemini-3.8-live. The 3.6 Flash page says: "Gemini 3.6 Flash will be retired on November 19, 2026. Upgrade to Gemini 3.8 Flash to avoid service disruptions." The Veo ids veo-3.1-fast-generate-001, veo-3.1-generate-001 print May 4, 2027 or later. That is a floor, not a confirmed shutdown on May 4, 2027. /shutdowns/google-vertex-2026-11-19/ lists the gemini-3.6-flash record. /shutdowns/google-vertex-2026-12-13/ lists the gemini-live-2.5-flash-native-audio record. A single-record date gets a page because the guide covers it.
Last reviewed . Publisher: SunsetDiff.
Steps
Read the two exact ids. The first table is every Google Vertex id in this guide whose shutdown date is 2026-11-19 or 2026-12-13. The replacement cell is the official replacement stored from the model versions page. Open the model page before you change a call.
Swap gemini-3.6-flash. The 3.8 Flash guide says: "Change model string in your API calls to gemini-3.8-flash." It also says: "Replace integer thinking_budget configurations with the string enum thinking_level". thinking_level="MINIMAL" is not available for 3.8 Flash. Explicitly setting thinking_level to MINIMAL will return an API validation error. Valid configurations are LOW, MEDIUM (default), and HIGH. The same page says temperature, top_k, and top_p are ignored. It says: "Passing frequency_penalty, presence_penalty or candidate_count will throw an active API error." Its checklist says: "Remove the following unsupported parameters: frequency_penalty, presence_penalty, candidate_count, temperature, top_p, and top_k."
Swap the Live model id. The Live migration page's samples set model to gemini-3.8-live. Affective dialog and proactive audio are always enabled on that model. In-session system instruction updates use send_context_update, or a context_update message. The 2.5 Live page lists regional locations. The 3.8 Live page lists us, eu, and us-central1. The migration page does not print a location field to change.
Test before each date. Call gemini-3.8-flash before November 19, 2026. Call gemini-3.8-live before December 13, 2026. The model versions page says retirement timelines may be extended, and they will not move earlier. The Veo rows say May 4, 2027 or later. Do not treat that cell as a shutdown on May 4, 2027.
What retires and when
These rows are the Google Vertex Gemini models in the dataset with shutdown date 2026-11-19 or 2026-12-13. The model id links to its SunsetDiff model page. Deprecation date, shutdown date, and replacement are the fields stored on that row. Not published means the field is empty. These rows do not store a deprecation_date.
The model versions page puts gemini-3.6-flash in the shorter-availability table: release date July 21, 2026, retirement date November 19, 2026, replacement model ID gemini-3.8-flash. It says: "Short-term availability models remain active until a replacement model is released and a retirement date is announced. When we schedule a model for retirement, we post a fixed date in the following table that gives you at least 45 days to migrate." And: "Even after a replacement launches, a model remains active until we announce its retirement date." The 3.6 Flash page says: "Gemini 3.6 Flash will be retired on November 19, 2026. Upgrade to Gemini 3.8 Flash to avoid service disruptions." The versions line on that page says retirement date November 19, 2026.
The same model versions page puts gemini-live-2.5-flash-native-audio in the 12-month table: release date December 12, 2025, retirement date December 13, 2026, replacement model ID gemini-3.8-live. The Live model page says retirement date December 13, 2026.
Both model pages say: "Listed retirement dates refer to retirement of support in Gemini Enterprise Agent Platform. Models may remain accessible through the Gemini API after these dates have passed." The Gemini API deprecations page is a different list. Checked on 2026-11-02, the shutdown cell for gemini-3.6-flash says "No shutdown date announced." It does not list November 19, 2026, December 13, 2026, or the id gemini-live-2.5-flash-native-audio. The stored Gemini API row for gemini-3.6-flash has shutdown kind none and no shutdown_date. See Gemini API model shutdowns on November 17, 2026.
gemini-3.7-flash is on the shorter-availability table with retirement date January 28, 2027 and replacement model ID gemini-3.8-flash. The stored shutdown date is 2027-01-28 (exact). January 28, 2027 is after mid-December, so that id is not in the table below. The page is /shutdowns/google-vertex-2027-01-28/. The January and March 2027 guide is Google Gemini 3.7 Flash and 2.5 Flash Image shutdowns.
2 Vertex Gemini ids with an exact shutdown date on 2026-11-19 or 2026-12-13.
The model versions page lists veo-3.1-fast-generate-001, veo-3.1-generate-001 with retirement date May 4, 2027 or later. See each row for its stored replacement. The dataset stores shutdown date 2027-05-04 and kind not_before.
The page says: "While retirement timelines may be extended, they won't be moved to an earlier date than what is listed." May 4, 2027 is the earliest day in that cell. It is not a confirmed shutdown on that day. This guide does not name a replacement the page does not print.
These Veo ids are not Gemini text or Live ids. A generateContent swap to gemini-3.8-flash does not replace a Veo video model.
2 Vertex Veo ids whose retirement cell says May 4, 2027 or later.
The 3.8 Flash guide says: "To migrate to 3.8 Flash from 3.7 Flash or 3.6 Flash, do the following steps." Then: "Change model string in your API calls to gemini-3.8-flash." Then: "Replace integer thinking_budget configurations with the string enum thinking_level". HIGH, MEDIUM, and LOW are the values it names. MEDIUM is the default. It also says: "thinking_level="MINIMAL" is not available for 3.8 Flash. Explicitly setting thinking_level to MINIMAL will return an API validation error. Valid configurations are LOW, MEDIUM (default), and HIGH."
Under mandatory API rules, the same page says deprecated sampling parameters (temperature, top_k, top_p): "These parameters are ignored by the backend." It says unsupported parameters: "Passing frequency_penalty, presence_penalty or candidate_count will throw an active API error." The checklist then says: "Remove the following unsupported parameters: frequency_penalty, presence_penalty, candidate_count, temperature, top_p, and top_k." The rules section and the checklist do not use the same verb for temperature, top_p, and top_k. This guide does not pick one. Remove the fields the checklist names, and do not send thinking_level MINIMAL.
The same page says: "History payloads cannot end with a "model" role turn." And: "Prefilled model turns are unsupported and must be removed when upgrading." It does not print an HTTP status or an error body for the retirement date itself. The model versions page says to update the application to the recommended upgrades, test mission critical features, and deploy the updates.
The quickstart below sets location to global. The 3.6 Flash page lists model availability as global, and multi-region us and eu. The quickstart is a generateContent call. It is not a Live session.
Retiring model id, same client shape as the 3.8 Flash quickstart
from google import genai
client = genai.Client(enterprise=True, project="PROJECT_ID", location="global")
response = client.models.generate_content(
model="gemini-3.6-flash",
contents="How does AI work?",
)
print(response.text)
3.8 Flash quickstart
from google import genai
client = genai.Client(enterprise=True, project="PROJECT_ID", location="global")
response = client.models.generate_content(
model="gemini-3.8-flash",
contents="How does AI work?",
)
print(response.text)
generateContent URL from the 3.8 Flash quickstart. The access-token header on that sample is left out.
curl -X POST \
-H "Content-Type: application/json" \
"https://aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/global/publishers/google/models/gemini-3.8-flash:generateContent" \
-d '{
"contents": {
"role": "USER",
"parts": { "text": "Why is the sky blue?" }
}
}'
The Live migration page describes the move from Gemini 2.5 Flash Live API Native Audio to Gemini 3.8 Live. Its samples set model to gemini-3.8-live. The retiring id on the model versions page is gemini-live-2.5-flash-native-audio. This guide does not invent a connect() call for that old id.
Affective dialog on 2.5 is "Supported through the enable_affective_dialog flag." On 3.8 Live the cell says: "Always enabled by default." Proactive audio on 2.5 is "Supported through the proactivity flag." On 3.8 Live that cell also says: "Always enabled by default."
For history seeding, the page says: "In Gemini 3.8 Live, if you want to use the send_client_content method to set up history seeding, in the HistoryConfig section of the setup message, set initial_history_in_client_content=True." For a system instruction change during a session, it says: "In Gemini 3.8 Live, call send_context_update in the Google Gen AI SDK, or send a context_update message over WebSockets, instead." It can replace the system instruction, the list of tools, or both.
The checklist says: "When the user is speaking, the scheduling=interrupt downgrades to scheduling=when_idle to provide a smoother user experience." The behavior table says that when the user is speaking, scheduling interrupt on 3.8 Live waits for the user to finish before playing the tool's response. On 2.5, that same cell plays the tool's response over the user. The page also says 3.8 Live supports blocking function calls when the declaration sets behavior=BLOCKING, and that a new input during that call cancels it.
The 2.5 Live page lists model availability in us-central1, us-east1, us-east4, us-east5, us-south1, us-west1, us-west4, europe-central2, europe-north1, europe-southwest1, europe-west1, europe-west4, and europe-west8. The 3.8 Live page lists multi-region us and eu, and United States us-central1. us-central1 is on both lists. The migration page does not print a location string to change. A location that is only on the 2.5 list is not on the 3.8 availability list this guide read.
The 3.8 Live page says retirement date September 25, 2027 or later. That is a floor for the replacement, not a December 13 date. The Live sample calls genai.Client() with no arguments. The 3.8 Flash quickstart uses enterprise=True, a project id, and location global. This guide does not merge those two clients.
Live connect sample from the Gemini 3.8 Live migration page
from google import genai
from google.genai import types
client = genai.Client()
config = types.LiveConnectConfig(
response_modalities=["AUDIO"],
media_resolution=types.MediaResolution.MEDIA_RESOLUTION_LOW,
)
async with client.aio.live.connect(
model="gemini-3.8-live",
config=config,
) as session:
...
Search the repository for the two Gemini ids in the first table, and for veo-3.1-fast-generate-001, veo-3.1-generate-001 if you use video models. A hit is a candidate. How to find every model ID in your codebase covers SDK argument names. The scan page builds a search from dataset ids that have a shutdown date.
gemini-3.6-flash on the Gemini API is a different provider row. The Vertex retirement date does not fill in a Gemini API shutdown date. Search configs that set a Vertex location as well as the model id. The Live pages do not share one location list.
Send one real generateContent request to gemini-3.8-flash before November 19, 2026. Send one real Live session to gemini-3.8-live before December 13, 2026. If a Live tool can return no rows, read that payload. The migration page says: "When a tool returns an uninformative or empty response, such as {"results": [], "total_available": 0} or None, the model can't distinguish between "permanently exhausted", "query window too narrow" or "invalid argument"."
veo-3.1-fast-generate-001, veo-3.1-generate-001 store 2027-05-04 with kind not_before. A floor is not an exact shutdown date, so there is no /shutdowns/google-vertex-2027-05-04/ page. Counting those 2 rows as exact would still be under 3.
Which Vertex Gemini models retire in November and December 2026?
The dataset lists gemini-3.6-flash, gemini-live-2.5-flash-native-audio. gemini-3.6-flash retires on 2026-11-19. gemini-live-2.5-flash-native-audio retires on 2026-12-13. The model versions page says: "While retirement timelines may be extended, they won't be moved to an earlier date than what is listed."
What is the recommended replacement?
The recommended replacement is the official replacement stored on each row. gemini-3.6-flash: gemini-3.8-flash. gemini-live-2.5-flash-native-audio: gemini-3.8-live. The 3.6 Flash page says: "Gemini 3.6 Flash will be retired on November 19, 2026. Upgrade to Gemini 3.8 Flash to avoid service disruptions."
What changes besides the gemini-3.6-flash model id?
The 3.8 Flash guide says: "Replace integer thinking_budget configurations with the string enum thinking_level". thinking_level="MINIMAL" is not available for 3.8 Flash. Explicitly setting thinking_level to MINIMAL will return an API validation error. Valid configurations are LOW, MEDIUM (default), and HIGH. It says: "These parameters are ignored by the backend." It also says: "Passing frequency_penalty, presence_penalty or candidate_count will throw an active API error." The checklist says: "Remove the following unsupported parameters: frequency_penalty, presence_penalty, candidate_count, temperature, top_p, and top_k." And: "Prefilled model turns are unsupported and must be removed when upgrading."
What changes for the Live API id?
Point the Live session at gemini-3.8-live. Affective dialog and proactive audio say: "Always enabled by default." The migration page says: "In Gemini 3.8 Live, call send_context_update in the Google Gen AI SDK, or send a context_update message over WebSockets, instead." And: "In Gemini 3.8 Live, if you want to use the send_client_content method to set up history seeding, in the HistoryConfig section of the setup message, set initial_history_in_client_content=True." And: "When the user is speaking, the scheduling=interrupt downgrades to scheduling=when_idle to provide a smoother user experience."
Do the Veo 3.1 models shut down on November 17, 2026?
No confirmed day is printed. The model versions page says May 4, 2027 or later, and the stored replacement is on each row. The page says: "While retirement timelines may be extended, they won't be moved to an earlier date than what is listed." The dataset kind is not_before. No Vertex Veo row stores shutdown date 2026-11-17.
Does the Gemini API use these same dates?
No. The Vertex pages say: "Listed retirement dates refer to retirement of support in Gemini Enterprise Agent Platform. Models may remain accessible through the Gemini API after these dates have passed." Checked on 2026-11-02, the Gemini API deprecations page lists gemini-3.6-flash with shutdown cell "No shutdown date announced." It does not list gemini-live-2.5-flash-native-audio. The November 17, 2026 Gemini API ids are on Gemini API model shutdowns on November 17, 2026.
Is there a shutdown page for November 19 or December 13?
A shutdown batch page is published when at least 3 model records share one provider and one exact shutdown date, or when a date is explicitly included. 2026-11-19 has 1 Google Vertex record with an exact shutdown date, so /shutdowns/google-vertex-2026-11-19/ exists. 2026-12-13 has 1 Google Vertex record with an exact shutdown date, so /shutdowns/google-vertex-2026-12-13/ exists. There is no /shutdowns/google-vertex-2027-05-04/ page because the Veo date is a floor.