Gemini 2.5 Pro and Flash calls change models on 1 December. What breaks, and what we'd do first

Google has deprecated gemini-2.5-pro and gemini-2.5-flash on the Gemini API. From 1 December 2026, any request that still names them gets answered by a newer model and billed at that model's price. Nobody has to deploy anything for that to happen. Find every app and API key that still calls the 2.5 models, test them on the replacement this month, and change the model string yourself before Google does it for you.

Start with quality. Your 2.5 Pro calls go to 3.8 Flash, a Flash model, and 2.5 Flash calls drop a tier to 3.5 Flash-Lite. Google's guides speak well of both, and they may well do better on your work. But nobody has run them on your prompts yet, and you don't want your users to be the ones who find out.

The second risk is outright failure. Google lists breaking API changes for both targets. They cover function calling and thought signatures, a new thinking setting, sampling parameters that no longer do anything, and prefilled model turns. A request that leans on one of those can return an error on 1 December even though nobody touched the code.

The third is the bill. Routed calls are charged at the new model's rates, and for 2.5 Pro users those rates change again on 1 January when Gemini 3.8 Flash comes off its introductory price. Google also says Gemini 3.8 Flash can use more tokens on long, complex tasks by design. So the per-token price is only half the story.

For engineers: Google does the routing on its side. Your application logs keep the name your code sent, so they won't show the switch. Google did something similar in October, when gemini-3.7-flash and gemini-3.5-flash dropped out of AI Studio and their endpoints started passing calls to newer models. For 2.5, Google lists 1 December as both the shutdown date and the routing date, so don't count on the old name answering for long. Move to the new name yourself.

Look at the gemini-2.5-flash row twice. Gemini 3.5 Flash-Lite starts at the minimal thinking level, and its guide warns that autonomous subagents with tool calls need medium or high "to prevent premature tool termination". An agent that kept calling tools on 2.5 Flash may stop early on the routed model, and that may not show up as an error. Google's guide calls Flash-Lite a strong migration path from 2.5 Flash and quotes better scores on reasoning and chart benchmarks. Plenty of extraction and classification jobs will be fine. Agent loops are the ones to test first.

The 2.5 Pro row has the opposite problem. Gemini 3.8 Flash defaults to medium thinking, takes smaller reasoning steps and checks its own work on multi-step goals, which is why Google warns it can use more tokens on long tasks. Thinking tokens are billed as output. If your Pro traffic is mostly short questions, set the level to low and measure.

Do the arithmetic for your own mix. Per token, a 2.5 Pro workload moving to Gemini 3.8 Flash pays 20% more for input and 25% less for output once standard pricing starts, and it pays less for very long prompts. Whether the total goes up depends on how many more tokens the new model spends on your tasks. Run your real prompts and read the usage numbers back from the API. Guessing won't get you close.

There are really only three choices, and doing nothing is one of them. Google's notice doesn't offer a way to stay on 2.5 Pro or 2.5 Flash past the date, so the decision is which newer model you pick, and whether you pick it or Google does.

Moving to Vertex AI is a fourth path for teams that want model access under Cloud IAM. It won't save you from this work. Vertex lists 20 October 2026 as the retirement date for the same models, so the model-level changes are similar. Pick it for governance reasons and plan it as its own project.

# example output
$ grep -rnE 'gemini-2\.5-(pro|flash)([^-]|$)' .
answer.py:14: MODEL = "gemini-2.5-flash"
agent.yaml:9: GEMINI_MODEL: gemini-2.5-pro
$ grep -rnE 'thinking_budget|temperature|top_' .
answer.py:22: temperature=0, thinking_budget=512
# trailing model turn = prefill, 400 on 3.6+
answer.py:31: role="model", text="Answer:"

Don't stop at the code search. Model names also sit in places grep can't see, like a low-code workflow or a notebook on someone's laptop. The AI Studio Usage page shows which keys in a project send traffic, and that's usually the quickest way to find a caller nobody remembers.

This notice came less than a month after the October one that rerouted 3.7 Flash and 3.5 Flash. It won't be the last. The durable fix is to treat the model name like a dependency version, and to treat the key behind it like a service account credential.

We made these examples up from Google's documented changes, so don't read them as Cohash clients or real incidents. You'll probably recognise one anyway.

Want a hand? Our AI practice can run this migration alongside your engineers, starting from your own traffic. We can give you a cost forecast at the new prices before anything moves. For Google Cloud work, see our Vertex AI page or hire a Vertex AI engineer. You can also start with an assessment.