Gemini 2.5 Pro and Flash get rerouted on December 1. What breaks, and what it costs

What changes on December 1

On November 2, 2026, Google’s Gemini API release notes deprecated two of its 2.5 models: “Effective December 1, 2026, all API requests to gemini-2.5-pro will automatically route to” gemini-3.8-flash, and requests to gemini-2.5-flash will route to gemini-3.5-flash-lite. Each is “billed at” the new model’s rates. Both old models have already left the picker in Google AI Studio.

So on December 1 the old name keeps answering, but another model does the work. That comes with 29 days’ notice, and it reverses what Google wrote on September 18, when it limited 2.5 access to people who had already used those models: “These models are not deprecated and will continue to be served until further notice through the API.”

gemini-2.5-flash-lite isn’t part of this. Google says it “is not deprecated and remains available through the API until further notice.”

You call	From December 1, served by	Price now, in / out per million tokens	Price of what serves it	Default thinking, before and after
gemini-2.5-pro	gemini-3.8-flash	$1.25 / $10, or $2.50 / $15 above 200,000 prompt tokens	$0.75 / $3.75 through December 31, then $1.50 / $7.50	Dynamic, then medium
gemini-2.5-flash	gemini-3.5-flash-lite	$0.30 / $2.50, with audio input at $1	$0.30 / $2.50, with audio input at $0.30	Dynamic, then minimal
gemini-2.5-flash-lite	Itself, with no date set	$0.10 / $0.40, with audio input at $0.30	No change	Off

Prices are from Google’s pricing page, checked November 4, 2026. On paper the limits don’t move: both old models and both new ones list 1,048,576 input tokens and 65,536 output tokens.

Who this reaches

Anyone whose requests name those two models on the Gemini API: your own code, a library, an automation, or any tool you’ve given a Gemini API key. Open-source projects are already finding it. primr and llm-exe opened issues on November 3 and 2 because their own model lists still had both as available.

Google’s notes cover API requests, so some products look unaffected:

• Antigravity. Its models page lists no Gemini 2.5 model.
• The Gemini app. Its model page names Flash-Lite, Flash, Pro and Argon, and doesn’t mention 2.5.
• GitHub Copilot. It dropped Gemini 2.5 Pro on July 31.

Google Cloud runs on its own schedule. Its lifecycle page gives all three 2.5 models, Flash-Lite included, a retirement date of October 20, 2026, which has passed. Its 2.5 Flash page says those dates refer to “retirement of support in Gemini Enterprise Agent Platform.” The November 2 notes don’t cover it, so don’t assume December 1 applies there.

What breaks

Google’s notes list the same changes for both moves: “Breaking API changes include function calling and thought_signature validation, thinking_level replacing thinking_budget, and removed sampling parameters and prefilled model turns.” In practice:

1. Function-call history. On Gemini 3 models, “you must pass back thought signatures during function calling, otherwise you will get a validation error,” a 400. On 2.5, returning them was optional. Google’s SDKs handle this if you “append the full model response object directly to history.” The risk is REST calls and code that rebuilds history in its own format, which is how an agent loop can end up failing partway through a task.
2. Prefilled model turns. “If the last non-empty turn in the request is a model turn, the API returns a 400 error.” Google suggests a system instruction or structured outputs instead.
3. Sampling settings. temperature, top_p and top_k are deprecated: “The API ignores these parameters and returns an error in future model generations.” So nothing fails yet, but a low temperature stops doing anything. candidate_count is “unsupported in Gemini 3 and later.”
4. Thinking settings. Replace thinking_budget with thinking_level. On generateContent, “Gemini 2.5 series models don’t support thinkingLevel,” so make that change together with the model string. A budget of 0 used to turn thinking off on 2.5 Flash, but Gemini 3 Flash and Flash-Lite “do not support full thinking-off.” On 3.8 Flash, minimal “will return an error.”
5. Function responses, on 3.8 Flash. Its checklist adds, “Only if using generateContent API: Ensure all FunctionResponse objects include call_id and name.”

The quiet one is Flash-Lite’s default. 2.5 Flash thinks dynamically unless told otherwise, while 3.5 Flash-Lite defaults to minimal. Google’s own advice for “autonomous subagents with tool calls, code execution, or multi-step reasoning” is to set medium or high “to prevent premature tool termination.” An agent that never set a level could get shallower work with no error at all.

What it costs

On 2.5 Flash, text and caching prices don’t change, and audio input gets cheaper. On 2.5 Pro, December is cheaper at $0.75 in and $3.75 out, but Google calls that an introductory price. On January 1, 2027 it becomes $1.50 and $7.50, so against 2.5 Pro on prompts up to 200,000 tokens, input costs 20% more and output 25% less. Gemini 3.6 Flash doubles on the same day.

Token counts matter as much as rates. Google says 3.8 Flash “can use more tokens on longer running and complex tasks, by design,” and output pricing includes thinking tokens. Measure a real task before assuming a saving.

Two other bills change:

• Google Search grounding. On the 2.5 models, paid projects get 1,500 free requests a day, then pay $35 per 1,000. On Gemini 3 and newer, they get 5,000 free a month, shared across those models, then pay $14 per 1,000, with a daily cap of 1,500. So the free allowance shrinks, and the price after it falls.
• Gemini 4 Argon. If you’d move a 2.5 Pro workload to Google’s flagship, its introductory price of $2 in and $10 out ends November 30. From December 1 it’s $4 and $20, and it isn’t on the free tier.

Whether Flash-Lite is good enough depends on the work. Google calls it a strong migration path from 2.5 Flash and reports higher scores on HLE (18.0% versus 11.0%) and CharXIV (74.5% versus 63.7%). Those are Google’s tests. In July, before 3.5 Flash-Lite came out, developers on Google’s forum said newer models trailed 2.5 Flash on their own benchmarks and latency, so run your own.

What to do before December 1

1. Find the names. Search your code and config. This skips gemini-2.5-flash-lite, and gemini-2.5-flash-image, which has its own date of March 15, 2027:

   grep -rnE 'gemini-2\.5-(pro|flash)($|[^-a-z0-9])' .

   Then check what grep can’t see: environment variables, databases, and the settings of tools holding your API key.

2. Choose the replacement yourself.

   • Extraction, routing or classification on 2.5 Flash: gemini-3.5-flash-lite, left at minimal.
   • Agents on 2.5 Flash: gemini-3.5-flash-lite at medium or high, or gemini-3.6-flash or gemini-3.8-flash if it stops early. Both cost more, especially from January.
   • 2.5 Pro: gemini-3.8-flash, Google’s named replacement, or gemini-4-argon for the hardest work.
   • Simple work already on gemini-2.5-flash-lite: it can stay, at the lowest price here. Access to 2.5 models is limited to people who already used them.
   • Leaving Gemini: some forum posters said they’d move to open models. See running a coding agent on a local model.

3. Switch now, while it’s your choice. The new names work today. Run real requests, fix any 400s, and compare the output and the token counts.

4. Handle saved conversations. A history started on 2.5 may not have the signatures Gemini 3 wants. For “transferring a trace from a different model,” Google allows a dummy signature, "skip_thought_signature_validator", in the thought signature field. It still calls injecting function calls “strongly discouraged.”

5. Or have an agent do the edit. Google offers a skill, installed with npx skills add google-gemini/gemini-skills --skill gemini-api-dev --global, and suggests running /gemini-api-dev migrate my app to Gemini 3.8 Flash. Review the result like any other diff.

What Google hasn’t said

• Which model a response came from. The notes don’t say whether a routed response names the model that served it, so log your own output quality and tokens around December 1.
• What routing does with old settings. Google says a lone thinkingBudget on Gemini 4 Argon maps to the nearest level with a warning. It doesn’t spell out the same for the Gemini 3 models these requests land on.
• Whose rate limits apply. Its rate limits page says limits vary by model, and the notes don’t say whether a routed request counts against the old model or the new one.
• How long routing lasts. The deprecations page lists December 1 as both models’ shutdown date. Its entries for gemini-3.7-flash and gemini-3.5-flash say “the endpoint is not shutting down,” and the 2.5 entries don’t. Treat the old names as temporary.
• When 2.5 Flash-Lite goes. “No shutdown date announced” is all the page says.

We’ll update this memo if Google moves the date or says how routed requests behave.