Google reroutes Gemini 2.5 Pro to 3.8 Flash on December 1. The bill can halve for 31 days, then input costs 20% more than 2.5 Pro did, and 2.5 Flash lands on a model that scores 12.

Three model strings, three different Decembers

Standard paid tier, input / output per 1M tokens, 2.5 Pro at its 200K-and-under rate. Gemini API changelog, deprecations and pricing pages, read November 6, 2026.

The November 2 changelog entry says requests "will automatically route" and bill at the new model's rates. We found no opt-out. The deprecations table also lists December 1 as the shutdown date, so treat the route as what the shutdown looks like, and don't expect a 404 to warn you.

This is a Gemini API change. Vertex AI had its own October 20 retirement, and the notice doesn't mention Vertex. Three speech models have an earlier date too: the 2.5 preview TTS models and the 2.5 Flash native audio preview shut down on November 17.

2.5 Pro gets 31 cheap days, then one ratio decides

December is easy: input falls 40% and output 62.5%. On January 1 the 3.8 Flash introductory price ends and every line doubles. That leaves input 20% above 2.5 Pro and output 25% below it.

Our arithmetic from Google's rates, with no cache or batch and the same token counts in every column, which is the assumption the next paragraph breaks.

Here's the check: open your 2.5 Pro bill and divide the input line by the output line. Above 1.25, January costs more than today on identical tokens. Cached reads move by the same 20% ($0.125 to $0.15), so count them as input. Prompts over 200K are the exception and stay well ahead, since 2.5 Pro charges $2.50 and $15 for those and 3.8 Flash has no long-context tier.

Identical tokens is a generous assumption. 3.8 Flash defaults to medium thinking, can't go below low, and bills thinking at the output rate. Google's launch post warned it might use more tokens. On that 30M/10M workload, 23% more output wipes out January's saving. December has far more room: output would have to triple on that same workload before you lost money.

One line does get much better, and few people will notice it. Context cache storage on 2.5 Pro costs $4.50 per 1M tokens per hour. On 3.8 Flash it's $0.50, then $1.00. A 500K-token cache held eight hours a day for 22 working days goes from $396 a month to $44, then $88.

The rate card flatters 3.8 Flash on agent work

If your workload is tool loops, the table above isn't your bill. Artificial Analysis prices a full agentic run per task. Bars below are that cost, with the models Google routes you to in the accent color.

Artificial Analysis Intelligence Index v5.0.1, read November 6, 2026. *Our estimate: today's figure doubled, because every line of the card doubles. Gemini 2.5 Pro and 2.5 Flash have no v5 score or cost, so we can't show the before. Don't compare these with scores from older index versions.

Run the list down and January stops looking like a Flash price. By our estimate the default 3.8 Flash setting will cost more per task than Gemini 4 Argon, which scores 20 points higher. Today's $6.26 is already about 6.5 times GPT-6.1 Terra at max effort, with Terra 21 points ahead.

Caching is a big part of the reason. 3.8 Flash on high and Gemini 4 Carbon land a cent apart ($7.77 and $7.76) despite the $4/$20 AA prices Carbon at. With no cache at all, AA's token counts would come to about $8,100 for 3.8 Flash on high and $34,600 for Carbon. AA's cost breakdown implies about 34% of 3.8 Flash's input tokens were cache hits, against about 91% for Carbon, so caching saved it 29% where Carbon saved 76%. Your agent may cache better. Check the cached token count in your usage metadata once you've tried it.

2.5 Flash keeps its price and swaps its model

Text-only, ungrounded users won't see this one on an invoice. 3.5 Flash-Lite has the same $0.30 and $2.50 text rate, and audio input drops from $1.00 to $0.30, a 70% cut. Voice pipelines get a real discount.

What you may notice is the output. Flash-Lite scores 12 on the index, 24th of the 29 models AA ranks with it, at $0.69 a task. We can't call that a drop from 2.5 Flash, which has no v5 score. We can say what else $0.69 buys: GPT-6 Luna on max scores 28 for $0.15, about a fifth of the money. Flash-Lite also starts at minimal thinking, and Google's migration notes suggest medium or high for subagents with tools. Those settings cost extra output tokens that this price match doesn't include.

Here's a 100M in, 10M out month, cheapest first. The oddity sits at the top of the list: the model Google didn't deprecate is the cheapest option in the family, at a quarter of the bill the route gives you.

Our arithmetic from each provider's pricing page, read November 6, 2026. No cache, no batch, no extra thinking. Access to 2.5 Flash-Lite is limited to users who have actively used the 2.5 models, and Google doesn't say whether 2.5 Flash history counts, so send a test call from your project now.

Search grounding: the cheaper rate comes with a ceiling

This is the line we'd check first, because it can cost more than the tokens and it can also start failing. Google's tool pricing changes in three ways when a request lands on a Gemini 3 model. The free allowance goes from 1,500 prompts a day to 5,000 requests a month, shared across every Gemini 3 and newer model. Overage falls from $35 to $14 per 1,000. And the Gemini API adds a default cap of 1,500 requests per day per project.

Our arithmetic for the paid tier, assuming one search per request. The 2.5 column assumes that model is the only one using its free pool; Flash and Flash-Lite share theirs. On the free tier, 2.5 Flash's 500 grounded prompts a day have no equivalent on 3.5 Flash-Lite, where the pricing page lists grounding as not available.

For a project with its own 1,500 free prompts, the lower rate never pays off under the default cap. Above about 170 a day you now pay for what was free, and that $560 tops a 2.5 Flash token bill of $55 ten times over. The $14 rate only wins above roughly 2,400 a day, which needs a quota increase. Google also bills Gemini 3 per search query, and one request can run several, so treat these as floors. It doesn't say whether the cap counts API calls or individual searches either, so you may hit it sooner. Above the cap you can request more quota in AI Studio, and Google's pricing notes also point high-volume users to Vertex AI, with a default of 1,000,000 search queries a day. Price that move before you make it, because we didn't.

What the router doesn't promise

Google lists breaking changes on this route: thinking_level replaces thinking_budget, sampling parameters and prefilled model turns are removed, and function calling gets stricter thought_signature validation. The 3.8 Flash guide adds candidate_count to the list, and the 3.6 notes say a prefilled turn returns HTTP 400.

Called directly, Google says sampling parameters are ignored and prefills return 400. We couldn't find anything that says the router behaves the same for a 2.5 request carrying the old fields, or what it does with thinking_budget. A silently dropped temperature or thinking_budget is the worst case, since nothing errors and only your eval scores and your bill change. Don't wait for December to learn which it is. Change the model string on a copy of your real payload today and look at what comes back.

Before November 30, by workload

Prompts over 200K on 2.5 Pro, or contexts where cache storage dominates the bill: take the route on purpose. Change the string yourself, set the thinking level rather than inheriting medium, and enjoy December. You're still ahead in January.

Short, input-heavy RAG on 2.5 Pro has the January problem. Run your evals against 3.8 Flash on low and against Terra, whose $1 and $5 are cheaper on both lines than 3.8 Flash in January. Agent loops should do the same test and add Argon, then compare cost per finished task rather than the cards. Argon's own price also rises to $4 and $20 on December 1, so budget from that number.

On 2.5 Flash, test 2.5 Flash-Lite and 3.5 Flash-Lite side by side. If 2.5 Flash-Lite passes, it costs about a quarter as much on that mix, with no shutdown date, though "no date" described 2.5 Flash too until this week. Voice work can simply take the route and keep the audio cut.

Anything grounded: pull your daily search count before anything else. Then put your own token counts through the calculator twice, once at $0.75 and once at $1.50.

Where each number came from

• Gemini API changelog - November 2 deprecations and routing, September 18 "not deprecated" note
• Gemini API deprecations - shutdown dates, limited access, 2.5 Flash-Lite status
• Gemini API pricing - Gemini rates except Carbon, cache storage, grounding allowances and cap, Vertex AI search quota
• What's new in Gemini 3.8 Flash and Gemini 3.6 - migration checklist, parameter changes, Flash-Lite thinking advice
• Artificial Analysis: 3.8 Flash medium, 3.5 Flash-Lite, 4 Argon, 4 Carbon, GPT-6.1 Terra, GPT-6 Luna, Haiku 5.5, DeepSeek V4.1 Flash
• OpenAI, Anthropic and DeepSeek pricing - competitor rate cards