Gemini 2.5 calls get rerouted on December 1. Pin the new model yourself before Google picks it for you
If your app, chatbot or automation still asks for gemini-2.5-pro or gemini-2.5-flash, a different model will answer it from December 1, at that model's price and with stricter rules. Some requests that work today will start failing with a 400 error.

From December 1, 2026, the Gemini Developer API answers requests for gemini-2.5-pro with gemini-3.8-flash, and requests for gemini-2.5-flash with gemini-3.5-flash-lite, billed at the new model's rates. gemini-2.5-flash-lite is unchanged. On the Agent Platform Gemini API (formerly Vertex AI), the 2.5 models already shut down on October 20. Find every call, remove frequencyPenalty, presencePenalty, candidateCount and prefilled model turns, move thinkingBudget to thinkingLevel, test on real cases, and set the model name yourself, ideally from Remote Config.
Your app still asks for 2.5. On December 1, a different model answers
Picture a support chatbot that went live last spring. It works, nobody has touched its settings since, and somewhere in the code it still says gemini-2.5-flash. On the morning of December 1 it keeps sending that name, and Google quietly hands the request to a different model.
That is the change Firebase announced in its November 2 release notes. Starting December 1, 2026, on the Gemini Developer API (the one behind Google AI Studio keys, and one of the two Gemini providers Firebase AI Logic can use):
- requests for
gemini-2.5-proare routed togemini-3.8-flash - requests for
gemini-2.5-flashare routed togemini-3.5-flash-lite gemini-2.5-flash-litestays as it is for now
If you use the Agent Platform Gemini API (formerly Vertex AI), this is old news in a worse form. All three 2.5 models shut down there on October 20, and Google's model page says a request to a shut-down model returns a 404 error. If anything of yours still calls them that way, it has been failing for weeks.
A reroute sounds gentler than a shutdown, and in one way it is: the request still gets an answer. But Google is clear that this can break things. The new model is billed at its own rates, it treats several settings differently, and some requests that work today will fail with a 400 error. A model swap you didn't choose is still a release. Test it like one.
Read one request the way December 1 will
The quickest way to see what changes is to take one request your product actually sends and go through it setting by setting. Here's a typical 2.5-era request, described in plain terms. It's illustrative, not taken from anyone's code.
- Model:
gemini-2.5-pro. From December 1 this is answered bygemini-3.8-flash, and you're billed at 3.8 Flash rates.gemini-2.5-flashis answered bygemini-3.5-flash-lite, at Flash-Lite rates. temperature: 0.2, plustopPandtopK. All of Google's general-use Gemini 3.x models ignore these. The request still works; the setting just stops doing anything. If you relied on a low temperature to keep answers steady, Google's advice is to write that rule into the system instructions.frequencyPenaltyorpresencePenalty. On a request routed to 3.8 Flash, either one makes the request fail with a 400 error.candidateCount. Same on 3.8 Flash: a 400 error. Google's migration guides call it unsupported in Gemini 3 and later, so take it out of the Flash requests too.- The last message in the conversation is a model turn you wrote yourself. Some apps prefill the start of the answer ("Here is the JSON:") to steer the format. Gemini 3.x models reject that with a 400 error. Use system instructions or structured output instead.
- Thinking set to
MINIMAL. 3.8 Flash doesn't support it; the request fails with a 400. Choose low, medium or high. thinkingBudget. Gemini 3.x models still accept a thinking budget for backwards compatibility, but Google strongly discourages it, especially from 3.8 Flash on, and recommendsthinkingLevelinstead.
Two of those seven quietly change behaviour. Four can stop the request outright. And the first changes the bill.
One more difference sits on the Flash side. According to Google's migration guide, 3.5 Flash-Lite defaults to the minimal thinking level, and Google suggests raising it to medium or high for anything that calls tools or works through several steps. If your 2.5 Flash feature runs a small agent, check this before you judge the new model's answers.
Decision guide
What each setting does after December 1
| Consideration | gemini-2.5-pro, routed to gemini-3.8-flash | gemini-2.5-flash, routed to gemini-3.5-flash-lite |
|---|---|---|
| Billing | 3.8 Flash rates: $0.75 input and $3.75 output per million tokens through Dec 31; $1.50 and $7.50 from Jan 1, 2027. | 3.5 Flash-Lite rates: $0.30 input and $2.50 output per million tokens. |
| temperature, topK, topP | Ignored. Put the rule in system instructions. | Ignored. Put the rule in system instructions. |
| frequencyPenalty, presencePenalty, candidateCount | The request fails with a 400 error. | candidateCount is unsupported in Gemini 3 and later. Remove all three and retest. |
| Prefilled final model turn | The request fails with a 400 error. | The request fails with a 400 error. |
| Thinking | MINIMAL fails with a 400. Use low, medium or high, and move thinkingBudget to thinkingLevel. | Defaults to minimal. Raise it for tool calls and multi-step work, and move thinkingBudget to thinkingLevel. |
A model swap you didn't choose is still a release. Test it like one.
Find every place that still says 2.5
The search is usually the hard part, because model names end up in more places than the main codebase. Here's where we'd look:
- App and server code. Search for
gemini-2.5across every repository, including Cloud Functions, scripts and anything a contractor left behind. - Firebase Remote Config and server prompt templates. If you followed Google's advice and keep the model name there, it's one place to change. Check that the value really is coming from there, and isn't also hardcoded as a fallback.
- Mobile apps already on people's phones. If an app version hardcodes
gemini-2.5-flash, every copy that hasn't been updated keeps sending it after December 1. A new release only fixes the copies people install. That's the strongest argument for moving the model name into Remote Config now, and making sure the version you ship reads it. - Automations. Low-code workflows, spreadsheet add-ons and internal tools where somebody typed a model name into a settings field a year ago.
- Shared API keys. If several apps or contractors use one key, list every one of them. Each is a place the old name can hide.
Write the list down as you go: where the call lives, which model it asks for, which settings it sends, and who owns it. That list becomes your test plan.
If you were also using -latest aliases, Google's model page doesn't recommend them even during development, because the model behind an alias can change with each release. We made the same argument in Do not promote what you cannot pin: if you can't say exactly which model answered, you can't explain a change in behaviour.
Test the routed model, then choose your own
You don't have to wait for December 1 to see what the routed model does. gemini-3.8-flash and gemini-3.5-flash-lite are available today. Take a few dozen real cases for each feature, remove the settings that will fail, and run them on the model your requests are headed for. Read the answers the way a customer would, not just whether the call returned 200.
Then decide whether that's the model you want. The reroute picks a destination for you, but nothing stops you from naming a different one.
- If you're on
gemini-2.5-flash, notice that you're being routed to the Flash-Lite tier. Google describes 3.5 Flash-Lite as a strong migration path from 2.5 Flash and publishes benchmark gains over it, but those are Google's numbers, not your use case. For simple extraction or classification it may be all you need. For a chatbot that has to reason, run the same cases ongemini-3.8-flashas well. - If you're on
gemini-2.5-pro, 3.8 Flash is the routed replacement. Firebase also lists Gemini 4 models among its recommended updates. Our Gemini 4 Hydrogen review covers the newest one and its price.
Price it while you test. On Google's pricing page, paid-tier 2.5 Pro costs $1.25 per million input tokens and $10 per million output tokens for prompts up to 200,000 tokens. 3.8 Flash costs $0.75 and $3.75 through December 31, then $1.50 and $7.50 from January 1, 2027. So for short prompts, January input is more expensive than 2.5 Pro was and output is cheaper. Whether your bill goes up or down depends on your mix of input and output, so price your own logs. 2.5 Flash and 3.5 Flash-Lite are both listed at $0.30 input and $2.50 output for text, image and video.
When a model passes, put its exact name in your code or Remote Config, and record why you picked it. Our note on scoring the job before you buy the model has a simple way to set the pass mark before you look at results.
Sequence of work
The dates that matter
- October 20, 2026
Gemini 2.5 Pro, Flash and Flash-Lite shut down on the Agent Platform Gemini API (formerly Vertex AI). Calls return 404.
- November 2, 2026
Firebase announces the December reroute and lists the breaking changes.
- Before December 1
Find the calls, fix the settings, test, and ship the model name you chose.
- December 1, 2026
Developer API requests for 2.5 Pro go to 3.8 Flash; 2.5 Flash goes to 3.5 Flash-Lite. 2.5 Flash-Lite is unchanged.
- January 1, 2027
Gemini 3.8 Flash paid-tier price rises from $0.75/$3.75 to $1.50/$7.50 per million input/output tokens.
Put this into practice
Here's a sensible order of work for the next two weeks:
- Today: search for
gemini-2.5everywhere: code, Remote Config, server prompt templates, automations and settings screens. Write down each call, its owner and its settings. - If you're on the Agent Platform Gemini API: check your logs for 404s from 2.5 model names. Those models shut down on October 20.
- This week: remove
frequencyPenalty,presencePenalty,candidateCountand any prefilled final model turn. SwapthinkingBudgetforthinkingLevel, and don't use minimal on 3.8 Flash. Move anything you used temperature for into system instructions. - Next: run real cases on the routed model, and on a second candidate if you were on 2.5 Flash. Price both against your logs, including the January price change on 3.8 Flash.
- Before December 1: ship the explicit model name you chose. For mobile apps, read it from Remote Config so the next change doesn't need an app release.
None of this is a big project for one feature. It gets bigger when nobody is sure how many features call Gemini, or which old app versions are still out there.
If you'd like help finding the calls or testing the replacement, that's our AI applications and agents work. When the model name lives inside an iOS or Android app, our mobile app development team can help with the Remote Config change and the release. Tell us what you're running.
Related work: Spin:Market
We build alongside the people doing the work.
Bring us the problem you’re thinking about.