November 19, 2026
Gemini 4 Neon Pricing: Budget Document Rewrites in 2026
A practical estimate for a document library, with separate rates for requests, revisions, and the work an editor still has to do.

The Gemini 4 Neon Standard API tier costs $3 per million input tokens and $15 per million output tokens. The output price includes thinking tokens. Google's Batch tier costs $1.50 per million input tokens and $7.50 per million output tokens. These are API rates, not the price of a Gemini app subscription.[4]
For a document rewrite project, estimate how many requests you'll make and allow for revisions. Sending a hundred documents as regular Standard requests still uses Standard pricing. Calling them a batch doesn't mean you're using Google's Batch service.
Google announced Neon on November 18, 2026. When we captured it on November 19, it was still on Techmeme and ranked eleventh on Hacker News, with 933 points and 724 comments.[1][6][7] That attention is deserved. But before you use "lower cost per task" to budget for finished writing, take a closer look at what the phrase means.
The Carbon comparison changes on December 1
Artificial Analysis measured Neon at $6.44 per benchmark task and Carbon at $7.76 using its standard, non-discounted pricing. The lab also found that Neon used about 15% more output tokens per task.[2] These results offer independent evidence about those evaluations, but they don't show what it would cost to rewrite your documents.
TNW's launch report distinguishes Carbon's temporary introductory price from its standard price.[3] That qualification matters when a headline calls Neon cheaper. Google lists Carbon Standard at $2 per million input tokens and $10 per million output tokens through November 30, 2026. On December 1, the rates rise to $4 and $20 per million tokens.[4]
Neon's token rates are higher than Carbon's introductory rates today, but lower than Carbon's scheduled December rates. That alone doesn't show which model will cost less for a given writing task. The billed token count and number of attempts still matter.
I wouldn't set an editorial budget using a benchmark's dollar figure. I'd choose a few representative documents, spell out the edits each needs, and track usage. A short announcement and a long research appendix shouldn't be lumped into one rough average.
A 100-document estimate you can actually inspect
This is a fictional example, not a measured result from Neon. A small organization wants to shorten 100 old help articles. Assume each first request is billed for 20,000 input tokens and 2,000 output tokens. The input count includes the document and instructions. The output count includes all billed output, including thinking, not just the draft text shown.
At Standard rates, input costs $0.06 and output costs $0.03, for a total of $0.09. The same pass on 100 documents costs $9. Four full passes with the same token counts cost $36. These totals don't include taxes, tools, storage, or other services, and assume no cache savings.[4]
At the published Batch rates, those token counts add up to $0.045 per request, or $4.50 for one pass through 100 documents.[4] That's only an arithmetic comparison. It doesn't show whether Batch will meet your deadline or fit your application. Check how your integration submits work before using Batch rates in a forecast.
Those assumptions no longer apply when the process changes. If a second request includes the original article, the first rewrite, and the conversation about it, count all of that as input. Don't use the first request's estimate for every later row. Longer model responses also use more output tokens.
Keep a brief record of every attempt, including the document ID, model and settings, service tier, billed input and output, cost, and whether the editor accepted it. Include rejected attempts. If a cheap first answer leads to several rewrites that get discarded, count them all toward the same document's total.
Google lists cached input at $0.15 per million tokens on Standard. Neon supports implicit caching, but not explicit caching.[4] I'd leave the discount out of the initial estimate and count any savings once they appear in the usage record. Reusing an attachment doesn't mean every input token got the cached rate.
Give a revision a job before paying for another
For the fictional help library, "make these articles better" is too vague as a batch brief. One editor might shorten the sentences, while another adds missing troubleshooting steps. Decide what each finished article should include before editing the folder.
Start with three very different documents: a short, ordinary page, a long page with exceptions, and an awkward page you know needs work. This pilot won't give you a reliable measure of accuracy. It's meant to catch a bad brief before you use it 100 times.
You could ask: "Shorten the introduction and remove repeated explanations. Keep the numbered procedure unchanged. Return the complete revised article, without an explanation of your edits." But treat this as something to check, not a guarantee that the model will follow it. Before you accept the result, compare the procedure with the original.
Then explain why you want another pass. "The introduction repeats paragraph two" gives the editor a clear task. "Try again" leaves them reviewing the whole article without knowing what to fix. If only the first paragraph needs work, ask for that paragraph alone, as long as it has enough context. Then read it in the full document to make sure it still fits.
Sometimes you should make the remaining edit yourself. Don't send a nearly finished page back just because another generation costs pennies. Someone has to compare the drafts and choose which one to publish. Track review time separately from API costs. Token savings don't show the full cost of the job.
Google describes better writing and document analysis in its Neon announcement.[1] That makes a limited trial worth considering, but it doesn't show that acceptance rates will improve. We haven't run this rewrite pilot, and this article isn't a hands-on review of the model.
Keep spoken editing notes separate from model purchases
When you read a long document, it can be easier to describe the edit you want: "The opening repeats the next paragraph. Keep the steps and shorten only the introduction." Add the note beside the draft, then review it. Decide whether to make the edit yourself or ask for a rewrite.
DictaFlow, which publishes this blog, offers desktop hold-to-talk dictation: hold the shortcut, speak, then release to insert text at the cursor.[5] It can capture an instruction in a text field, but it can't choose a Gemini API tier, submit a Batch job, or check the rewrite. This article doesn't announce a Neon integration.
If you mainly want to dictate your own words in apps you already use, start with dictation. If you manage a document library, consider testing Neon in a small API pilot. Track usage and set clear acceptance rules. For the dictation setup, see the desktop instructions rather than the model's token-price table.
Keep the pilot's originals, approved revisions, and usage records together. Use them to price the remaining documents. Before approving a December run, check the live Gemini 4 Neon pricing page again and date your estimate. Record the rates and revision counts you used to build the budget.
Disclosure: Published by DictaFlow with AI-assisted research and drafting. The help-library project and editing instructions are hypothetical.
Sources
[1] Google: Introducing Gemini 4 Neon
[2] Artificial Analysis: Gemini 4 Neon intelligence and task costs
[3] TNW: Gemini 4 Neon launch coverage
[5] DictaFlow Consumer product facts