Doing Nothing About Gemini 2.5 Pro Looks Smart Until January

The discount hiding in a shutdown notice

On December 1, every request to gemini-2.5-pro stops reaching Gemini 2.5 Pro. Google's deprecations page says the calls will route to Gemini 3.8 Flash and bill at 3.8 Flash rates. You don't even have to change the model name.

I read that and thought it was free money. Input drops 40%. Output drops 62.5%. You get a newer model and a smaller bill for doing nothing at all.

Then I opened the 3.8 Flash release notes. Those prices are introductory and they end on December 31. So your first routed month is the cheapest one on Google's schedule.

Look at the input column. From January, 3.8 Flash charges more than the model it replaced, for every input token and every cached token. Only the output rate is still lower. Whether you come out ahead depends on one ratio, and it takes about ten seconds to check.

Run the January number, not the December one

Here are four made-up workloads priced at list rates. Swap in your own volumes. The tokens are per month, and the first three stay under the 200K prompt line.

The document workload is the one that should worry you. It reads a lot and writes a little, which describes a great many RAG and extraction jobs. It saves $375 in December and then pays $75 a month more than it does today.

The rule is simple. Divide your monthly output tokens by your input tokens. Under 10%, you pay more from January than you pay now. At 10%, you tie. Above that, you save. That assumes your token counts stay the same, and they might not.

Very long prompts are the clear winners. Today a prompt over 200K tokens bills the whole request at 2.5 Pro's higher rate. Our 3.8 Flash record has no such long-context tier, so the contract review job still pays 41% less in January than it does now.

Want your own number without a spreadsheet? The free model switch calculator takes your tokens and shows the scheduled rate next to today's. It's a quick way to learn which side of that 10% line you're on before Google's invoice tells you.

The sentence that can eat the rest of your savings

Google's own guide says 3.8 Flash "can use more tokens on longer running and complex tasks, by design." The table above can't see that, because it assumes the same token count on both models.

Thinking is billed with your output. The thinking docs say you pay for the full thoughts, even though the API only hands back a summary. 3.8 Flash thinks at medium by default. 2.5 Pro also bills its thinking as output, so what matters is how much total billed output changes when the model does.

I didn't run a benchmark for this post, so I can't give you that multiple, and it will differ for every prompt set. What I can give you is how much room you have. This is how many times today's billed output 3.8 Flash can produce before your bill catches up with what you pay now.

Read the January column again. The chat assistant has 21% of headroom. If 3.8 Flash bills 21% more output on the same conversations, your January saving is gone. If it bills twice as much, that $1,175 month turns into $1,650, or $5,700 a year more than you pay today. The December invoice won't warn you, because at December rates the same usage still looks like a saving.

That's the part I'd worry about. You get one cheap month, you assume the redirect worked out, and the higher rate arrives quietly on January 1.

What stops working, and what fails without telling you

Google calls these breaking API changes, and they're in the migration guide and the Gemini 3 API changes. The silent ones scare me more than the errors.

• temperature, top_p and top_k are ignored. No error, no warning. If you set temperature to 0 for an extraction job, that setting stops doing anything. Google's suggestion is explicit rules in a system instruction.
• Prefilled model turns return a 400. If you end a request with a model turn to force JSON or skip the preamble, that trick now fails. Use structured outputs or a system instruction.
• thinking_budget has to become thinking_level. Your choices on 3.8 Flash are low, medium and high. Sending minimal returns an error.
• candidate_count is unsupported in Gemini 3 and later.
• On generateContent, every FunctionResponse needs a call_id and a name. Google also says to check tool calls for Malformed_Function_Call errors.

One thing I couldn't pin down: these rules are written for calls to the new models, and Google doesn't spell out how a routed gemini-2.5-pro request treats each old field. It also doesn't say routed requests are exempt. Assume they apply and test it, because I'd rather find a 400 in staging in November than in production on December 1.

If you're on Gemini 2.5 Flash

You get the easier deal. gemini-2.5-flash routes to Gemini 3.5 Flash-Lite, and the text list price is the same on both: $0.30 input, $0.03 cached input and $2.50 output. That chat workload costs $290.00 before and after, and 3.5 Flash-Lite has no scheduled increase on Google's page as of our last check.

The catch runs the other way. Flash-Lite's migration notes set thinking_level to minimal by default, while older Gemini models pick their thinking level on the fly. That could trim your output tokens, but Google warns that autonomous agents need medium or high to avoid stopping tools too early. So a cheap agent can get cheaper and worse in the same week, and the bill won't show you the second half. The sampling and prefill rules above apply here too. gemini-2.5-flash-lite isn't part of this change.

Four ways to hang on to the December price

You can't stop the rates going up, but you don't have to pay all of it. Here's the chat workload from January onward.

• Use Batch for anything that can wait. Batch stays at half the Standard rate, so from January it costs $0.75 and $3.75, which is exactly December's Standard price. Overnight reports, backfills and evals all qualify.
• Set thinking_level to low on everyday tasks. Google itself says not every workflow needs this much checking and that lowering the effort cuts token use. Compare quality first, then keep the saving.
• Send easy jobs to Flash-Lite. Moving this whole workload from 3.8 Flash saves $9,120 a year at list rates.
• Cache your repeated prompt. From January, cached input on 3.8 Flash is $0.15 against $1.50 fresh, so each hit is 90% off. This is the best fix for the document workload.

Before you move hard work to Flash-Lite, read what happened when I graded it. It got every everyday task right within three attempts and still missed 6 of 270 hard items. That result could save you a bad rollout, and it cost you nothing to read. To see how much of that 90% cache discount you'd keep, try the free prompt caching calculator with your real hit rate. If Gemini's new prices are the deal-breaker, this list of same-price models gives you a shortlist to test.

What I'd do before December 1

• Pull last month's input, cached and output tokens for each Gemini model. Divide output by input to see which side of the 10% line you're on.
• Search your code for temperature, top_p, top_k, candidate_count, thinking_budget and any request that ends on a model turn.
• Send 100 real requests straight to gemini-3.8-flash at the default setting, then again at low. Compare the answers and total_thought_tokens with what 2.5 Pro gives you today.
• Multiply the new token counts by the January rates. December's number is the one that will mislead you.
• Decide what goes to Batch, what goes to Flash-Lite and what stays on Standard, and set a spend alert for January.

All of this covers the Gemini API, checked on November 7, 2026. Vertex AI has its own lifecycle pages, so check those separately. Upcoming dates are in our model price monitor, and current rates are on the Gemini 3.8 Flash and Gemini 3.5 Flash-Lite pages. Put your workload in the calculator while you still have time to act on the answer. It's stored in your browser only and nothing alerts you, so open it again in January.

FAQ