From Google's Gemini API pages, the Vertex AI model-versions page, a Cursor forum thread and library source code, 17 November 2026 · Updated

gemini-2.5-flash and gemini-2.5-pro deprecation: requests route to other models from 1 December 2026

On the Gemini API, Google says that from 1 December 2026 it sends every gemini-2.5-flash request to gemini-3.5-flash-lite automatically, and every gemini-2.5-pro request to gemini-3.8-flash, billed at the new model's rates. Google lists breaking API changes for both new models. Its table puts 1 December 2026 in the "Shutdown date" column for both IDs, and, unlike its rows for gemini-3.7-flash and gemini-3.5-flash, it does not say the endpoint stays up. We do not know what a request to these IDs returns from 1 December. We read Google's pages on 17 November 2026; this page quotes them and does not add to them. We did not call the API.

What Google's page says

What the Gemini API deprecations page says, read 17 November 2026 (UTC) at 13:13. This is in our words, except the parts in quotation marks:

  • gemini-2.5-pro is deprecated and can no longer be selected in Google AI Studio. Through the API, Google keeps it open only to accounts that already called it. From 1 December 2026, all API requests to it "will automatically route to gemini-3.8-flash", billed at Gemini 3.8 Flash rates. Before then, Google asks you to read the Gemini 3.8 Flash migration guide, which lists the API changes that can break your code.
  • gemini-2.5-flash is deprecated and can no longer be selected in Google AI Studio. From 1 December 2026, all API requests to it "will automatically route to gemini-3.5-flash-lite", billed at Gemini 3.5 Flash-Lite rates. Before then, Google asks you to read the Gemini 3.5 Flash-Lite migration guide, which lists the API changes that can break your code. Google keeps API access to both 2.5 models for accounts that already called them.
  • gemini-2.5-flash-lite is not deprecated, and Google says it "remains available through the API until further notice".
Model IDRelease dateShutdown date on the pageReplacement on the page
gemini-2.5-pro17 June 20251 December 2026gemini-3.8-flash
gemini-2.5-flash17 June 20251 December 2026gemini-3.5-flash-lite
gemini-2.5-flash-lite22 July 2025"No shutdown date announced"none listed

The page says the dates in its table are "the earliest possible dates on which a model might be retired", and that Google "will communicate the exact shutdown date to users with advance notice". It defines a model that is shut down as one whose "endpoint is no longer available". For these two IDs it gives the routing sentence and the table date, and says nothing more about what happens to a request from 1 December. We do not know.

gemini-2.5-flash-lite is a different ID. Google's page names no routing for it and says it "remains available through the API until further notice".

What Google lists as breaking

Google's release notes of 2 November 2026 name the same breaking areas for both models: checks on function calls and on thought_signature, a new thinking_level setting in place of thinking_budget, sampling parameters that are gone, and an end to prefilled model turns. The two migration guides give the detail. Below is what they say, and a pattern to search for in your code. A search hit means "look here", not "this breaks".

ChangeIn which guideWhat the guide saysSearch your code for
Thinking settingBoth (release note: "thinking_level replacing thinking_budget")3.8 Flash guide: replace thinking_budget with the string enum thinking_level. On 3.8 Flash, minimal is not supported; choose low, medium (the default) or high. 3.5 Flash-Lite: its default is "minimal", which the guide recommends for high-volume extraction, routing or classification. For agents that call tools, run code or reason over several steps, it says to set "medium" or "high" so that tool use does not end too earlythinking_budget, thinkingBudget
Sampling parametersBothtemperature, top_p and top_k are deprecated, and the guide says to remove them from all requests. Google's Gemini 3.6 page says they are "deprecated and ignored", and that in future model generations supplying them returns an HTTP 400 errortemperature, top_p, topP, top_k, topK
Candidate countBothcandidate_count is "unsupported in Gemini 3 and later"candidate_count, candidateCount
Prefilled model turnsBothA request whose last non-empty turn comes from the model role gets a 400 error back.a contents list whose last item has role: "model" (no reliable search; read your request builders)
Function callingBoth, except the last itemMultimodal assets go inside the response payload, and inline instructions are formatted with \n\n; for Malformed_Function_Call errors tied to pre-tool text, see Google's workarounds page. 3.8 Flash only, and only on the generateContent API: every FunctionResponse object needs both call_id and nameFunctionResponse, function_response, functionResponse
Multi-turn state3.8 Flash; thought signatures both"Standardize multi-turn conversations on server-side previous_interaction_id"; thought signatures must be preserved (Google's "Thought signatures" page)previous_interaction_id, thought_signature, thoughtSignature

The 3.5 Flash-Lite migration guide is on Google's Gemini 3.6 page, and the 3.8 Flash guide points to that page for the parameter rules. The 3.6 page lists API changes that cover Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and every Gemini model Google releases after them. Those changes say the sampling parameters are "deprecated and ignored", and that supplying them "In future model generations" returns an HTTP 400 error; the page does not say which generation that is. We did not check whether a 2.5 request that sets these fields is changed by the reroute before it reaches the new model. The page does not say, and we could not test it.

What a breaking change returned on a new model

No rerouted call has failed that we know of, and we have no API key, so we could not make one fail. One public report shows a change from the list on a 3.x model, and it is not a prediction for the reroute:

  • A Cursor forum thread, 17 August 2026, titled "Custom Gemini Model Not Supported": a user added gemini-3.5-flash-lite as a custom model and ran a query with a tool call. A Cursor reply of 20 August says that on later turns with tool calls "Google returns 400 Function call is missing a thought_signature", because the client did not resend thought signatures. That was a third-party client's bug, not a statement from Google.

We found no public report of an error from thinking_budget, sampling parameters or candidate_count on gemini-3.8-flash or gemini-3.5-flash-lite.

Find these IDs in your code

This searches every file in a folder that grep reads as text for the two exact IDs. That includes .env files, Dockerfiles, notebooks, shell scripts, Terraform files 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; if your virtual environment has another name, add one more --exclude-dir= for it. It does not match gemini-2.5-flash-lite, gemini-2.5-flash-image, gemini-2.5-flash-preview-*, gemini-2.5-pro-preview-*, -tts or -native-audio IDs, which are different models.

grep -rnEI 'gemini-2\.5-(flash|pro)([^A-Za-z0-9_-]|$)' . \
  --exclude-dir=node_modules --exclude-dir=.git --exclude-dir=.venv --exclude-dir=venv

The pattern also matches prefixed forms such as models/gemini-2.5-pro and vertex_ai/gemini-2.5-flash. It finds nothing in a repository that sets no model name; see the next section. A hit inside a minified JavaScript file prints the whole line, which can be very long; add | cut -c1-200 to the end to shorten each line.

grep reads a file saved as UTF-16 as binary and skips it, so none of the searches on this page list a line from one. If some of your files are UTF-16, convert each one to UTF-8 and search the output, for example iconv -f UTF-16 -t UTF-8 config.ps1 | grep -nE 'gemini-2\.5-(flash|pro)([^A-Za-z0-9_-]|$)'. We tested that on a UTF-16 file that starts with a byte-order mark.

To look for the settings, function-calling and state names in the table, search for every name in its right column (the prefilled-turn row has no reliable search). This matches any file that uses them, Gemini or not:

grep -rnEI -e 'thinking_budget|thinkingBudget|temperature|top_p|topP|top_k|topK|candidate_count|candidateCount' \
  -e 'FunctionResponse|function_response|functionResponse|previous_interaction_id|thought_signature|thoughtSignature' . \
  --exclude-dir=node_modules --exclude-dir=.git --exclude-dir=.venv --exclude-dir=venv

We ran both blocks on a test app folder in bash with GNU grep 3.12 and in zsh 5.9 with GNU grep 3.11. The first listed all 17 lines that name one of the two IDs in the app's own files, including a notebook, a Dockerfile, a Terraform file, a shell script and .mjs and .cjs files. It listed nothing from the app's installed packages, a binary file or a UTF-16 file, and no line naming -lite, -image, -preview-*, -tts or -native-audio. The second listed the 8 lines, in Python and TypeScript files, that use those names. In a folder that names only those other IDs and the replacements, both printed nothing. We also ran the first block's command with BSD grep 2.6.0 in zsh on a second test folder, where it listed all 15 of that folder's own files that name an ID.

Correction, 18 November 2026: an earlier version of this page said these searches read every text file, and its second search covered only the thinking, sampling and candidate-count names. grep skips UTF-16 files, as the paragraph above the second search now says, and the second search now also covers the function-calling and state names in the table.

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-flash and a pasted TypeScript file calling gemini-2.5-pro each returned the 1 December routing and the replacement for the Gemini API. A gemini-2.5-flash-lite line is reported under the Gemini API as "No shutdown date on the page". The check's summary line at the top also counts the Vertex AI dates in the section below, so for that same line it can report "1 already shut down"; read the Gemini API section if that is the API you call.

Libraries that choose the model for you

If you pass no model, a library's default decides. We read the latest release of each library below, downloaded without installing, on 17 November 2026. We did not run any of them.

Library (latest read)Default model with no model passedUses gemini-2.5-flash or gemini-2.5-pro by default?
google-adk 2.14.1 (12 Nov)LlmAgent default gemini-3.5-flash. The Gemini class itself has model: str = 'gemini-2.5-flash'. Several evaluation, optimisation and user-simulation helpers default to gemini-2.5-flash; the agent builder assistant defaults to gemini-2.5-proYes, in the Gemini class and in those helpers
livekit-plugins-google 1.8.12 (Python)LLM gemini-3.8-flashNo
@livekit/agents-plugin-google 1.10.0 (JS)LLM gemini-2.0-flash-001No, and see the note below
pipecat-ai 1.16.0GoogleLLMService gemini-3.6-flash; Vertex gemini-3.8-flashNo
agno 3.2.0Gemini id gemini-3.7-flashNo
langchain-google-genai 4.5.0none: the model is a required argument; its examples use gemini-3.5-flash and gemini-3.1-pro-previewNo
llama-index-llms-google-genai 0.12.0DEFAULT_MODEL = "gemini-3.7-flash"No, and see the note below
pydantic-ai-slim 2.69.0none found: the caller passes the modelNo
@ai-sdk/google 4.0.121, @genkit-ai/google-genai 1.44.0none found: the caller passes the model; both list gemini-2.5-flash and gemini-2.5-pro among known IDsNo
@mastra/core 1.83.0 (memory code in @mastra/memory 1.41.1)Agents: none found, the caller passes the model. Observational memory, which is off unless you set observationalMemory: with no memory model set it picks google/gemini-2.5-flash when GOOGLE_GENERATIVE_AI_API_KEY is set (unless you pass canUseProvider or autoModels), and in some setups without that key; a memory model set to "default" always picks it. Its defaults also send a Google thinkingBudget (215 for the observer, 1024 for the reflector) unless you set providerOptionsYes, in observational memory
crewai 1.15.41Gemini provider gemini-2.0-flash-001; lists both 2.5 IDs among known modelsNo, and see the note below

For google-adk we also read releases 2.11.0, 2.12.0, 2.13.0, 2.14.0 and 2.14.1 (all since 1 October): the Gemini class default is gemini-2.5-flash in every one, and the LlmAgent default is gemini-3.5-flash in every one. For the other libraries we read only the latest release, so we do not know whether an earlier release since 1 October changed a default. "None found" means we searched every file in the package for a default Gemini model ID and found none.

Separate findings, not part of the 1 December change: the JavaScript LiveKit plugin and crewai's Gemini provider default to gemini-2.0-flash-001, which Google's page lists with a shutdown date of 1 June 2026. llama-index's Google GenAI default is gemini-3.7-flash; Google's page says that endpoint "is not shutting down and automatically routes API requests to gemini-3.8-flash".

Apart from @mastra/core's observational memory (in the table), we did not check whether any of these libraries sends thinking_budget, a sampling parameter or candidate_count on its own when you pass none.

Vertex AI says something different

Google Cloud's Vertex AI model-versions page, read 17 November 2026, lists gemini-2.5-pro, gemini-2.5-flash and gemini-2.5-flash-lite with a retirement date of 20 October 2026, which has passed. It names gemini-3.8-flash or gemini-3.5-flash as the replacement for Pro, gemini-3.8-flash or gemini-3.5-flash-lite for Flash, and gemini-3.8-flash, gemini-3.5-flash-lite or Gemma 4 for Flash-Lite. The page says "retirement timelines may be extended, they won't be moved to an earlier date than what is listed". It does not describe a routing step like the one on the Gemini API. We do not know what a request to these IDs on Vertex AI returns today. Note that the Gemini API page says gemini-2.5-flash-lite has no shutdown date; Vertex lists a date for it.

What to do

  1. Search your code for the exact IDs, and for the names in the table.
  2. Pick the model you want. Google's replacement for gemini-2.5-pro is gemini-3.8-flash, and for gemini-2.5-flash it is gemini-3.5-flash-lite. The Gemini 3.6 page calls 3.5 Flash-Lite a "Strong migration path from Gemini 2.5 Flash" and gives benchmark figures, which we do not reproduce. It is a different model, not the same one under a new name.
  3. Google's release note says: "Before December 1, 2026, update your model string to" the replacement, and review its migration guide. Setting the name yourself also means you choose the model, rather than the reroute choosing it. Then apply the changes in the table for that model. If you move from gemini-2.5-flash to gemini-3.5-flash-lite and your code calls tools, read the thinking row first.
  4. If you use google-adk, search for Gemini( and open each hit to see whether the call passes model=; a call can span several lines. A call without one, such as Gemini(use_interactions_api=True) in google-adk's own code examples, uses gemini-2.5-flash. Also check the evaluation, optimisation, simulation and agent-builder settings the table names, which default to a 2.5 ID when you set no model.
    grep -rnI 'Gemini(' . --include='*.py' \
      --exclude-dir=.git --exclude-dir=.venv --exclude-dir=venv
  5. If you use @mastra/core, search for observationalMemory. Where it is on and sets none of model, observation.model or reflection.model, or sets one to "auto" or "default", its memory calls can use gemini-2.5-flash. Set a model there, and check providerOptions, which otherwise carries a Google thinkingBudget.

What we did not check

  • Any Gemini API call. We have no API key, so we do not know what a rerouted request returns, or whether a request that sets temperature or thinking_budget is rejected, ignored or changed.
  • Whether the Gemini API behaves as Google's page says on 1 December. We read the page on 17 November.
  • Vertex AI behaviour. We read only the model-versions page.
  • Library releases other than the latest for each library, apart from google-adk.
  • PowerShell, Windows, and macOS itself. We ran BSD grep 2.6.0, not a Mac.
  • Text encodings other than UTF-8, apart from one UTF-16 file with a byte-order mark.
  • Price. Google sets the rates; see its pricing page.

Sources

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.

Check your code for these model IDs and where they route