Teams leaving Nano Banana 2 by October 29 find Google's docs disagree on the 4K price

The person who owns the image budget has two Google tabs open. The Gemini API pricing page puts a 4K Nano Banana 2.1 image at 2,520 output tokens, or $0.0756 at $30 per million ^[6][5]. The Vertex AI model page lists the same 4K image at 3,780 tokens. At the same rate, reAPI calculates, that is about $0.1134 ^[18].

The Vertex count is 50% higher, a gap of $0.0378 on every 4K image ^[20]. On 100,000 4K images a month, the gap is $3,780 ^[21]. That is more than the $3,340 a month that 100,000 1K images save by moving from Nano Banana 2 to 2.1 at list prices ^[17].

Teams calling Google directly cannot wait this out. Google's changelog marks Nano Banana 2, model id gemini-3.1-flash-image, as deprecated and tells developers to migrate to 2.1 ^[3]. reAPI's pricing guide, which compiles Google's list prices as of October 7, went up 22 days before the shutdown ^[4][16]. Its advice for 4K-heavy workloads is to "check your first invoice rather than budgeting from either page" ^[19]. reAPI also resells the model at one flat price per image, so the guide has a product to sell ^[12].

The pitch is half price. The halving applies to the image tokens: Google cut the image output rate from $60 to $30 per million ^[8]. A production request can carry other lines. Each reference image is 1,120 input tokens on Vertex, about $0.0017 at $1.50 per million, and a 14-reference fusion job adds about $0.024 ^[10]. Thinking is billed separately and defaults to medium ^[1]. In three reAPI requests with different prompts, thinking used no tokens at minimal, 892 at the default and 2,658 at high ^[2]. The high figure costs about $0.02 ^[13].

Add one grounding query past the free monthly allowance, at least $0.014 ^[11], and a 1K image with high thinking comes to about $0.0675 ^[14]. Nano Banana 2's 1K list price was $0.067 ^[7]. The guide does not say what thinking or grounding added on the old model, so the two figures are not like for like.

I'd migrate now and budget 4K at the Vertex count of 3,780 tokens until the first invoice shows which number Google bills. The cost of that choice is a forecast that runs up to $0.0378 per 4K image high if the Gemini API page turns out to be right ^[20]. Testing costs money from the first call, because the free tier is not available for this model ^[9]. Workloads that can wait for results get the Batch tier at half the Standard price ^[5].

Two measurements sort the rest: the share of output at 4K, and whether requests carry grounding or high thinking. A team rendering mostly 1K and 2K with plain requests will find the list price close to the bill, and a saving close to half ^[7]. Switch on high thinking and one grounding query at 1K, and the extras, about $0.034, outweigh the $0.0336 image line ^[15]. That team should log thinking tokens and search queries per request in the first week, since one request can trigger more than one search ^[11]. A 4K-heavy team with plain requests can budget at $0.1134 and correct down once billed ^[18]. For 4K-heavy work with extras on, a flat per-image price that covers resolution, references, grounding and thinking level, like the one reAPI posts on its model page, is the comparison to run against Google's metered total ^[12].