Nano Banana 2.1 API Pricing: $0.0336 per 1K Image, Half Price in Batch Mode

Bottom line: Google’s Gemini API lists Nano Banana 2.1 at $0.0336 per 1K image (vs. $0.067 for Nano Banana 2), with batch mode at $0.0168. CNET’s October 12 roundup still names Nano Banana 2 as its best overall image generator.

The sticker and the batch column are not the same invoice. The Next Web’s October 9 account says Google released Nano Banana 2.1 on October 6, built on Gemini 3.6 Flash, and that the Gemini API price list puts a 1K image at $0.0336, against $0.067 for Nano Banana 2. A 4K image is $0.113, against $0.151. Powerdrill’s guide, which says it read Google’s pages on October 13, prints the same $0.0336 and the same $0.113, and then prints a second column. Batch is $0.0168 for 1K and $0.0567 for 4K. Katelyn Chedraoui’s CNET roundup, updated October 12, 2026, 12:41 pm EDT, still names Nano Banana 2 as the best overall image generator. Six days after the GA note Powerdrill quotes, the roundup’s top pick is the model the deprecation page is pointing away from.

We already had a per-second price that depended on which card you read. Vidu Q4 Preview was $0.014 in one place and $0.045 in another. We already had a model whose public call was not the call. Gemini 4 Argon was the cyber-partner door first. We already had a small-model sticker that was not an average saving. Haiku 5.5’s $0.10 line was that cut. This week’s image model has a standard column, a batch column, and a review that has not changed names.

Nano Banana 2.1’s model id is on one page. The 3.6 Flash line is on another.

Powerdrill quotes the Gemini API release notes for October 6, 2026: Gemini Nano Banana 2.1 generally available, model id gemini-nano-banana-2.1, “the latest high-efficiency image generation and conversational editing model.” The notes, as Powerdrill renders them, call it an update, not a new family, and say it maintains Flash-level speed and cost efficiency. They do not, in that rendering, print “Gemini 3.6 Flash.” The Next Web does. It says Nano Banana 2.1 is a new image model built on Gemini 3.6 Flash. One page has the id. The other has the base-model name. A client config that needs the string should copy gemini-nano-banana-2.1. A slide that needs the base should cite The Next Web, not the release-notes paragraph that did not include it.

The deprecation line is also a name problem. Powerdrill says the notes deprecate gemini-3.1-flash-image, with no shutdown date, and tell developers to migrate to gemini-nano-banana-2.1. The Next Web says Google’s deprecations page lists Nano Banana 2.1 as the replacement for Nano Banana 2, and that no shutdown date has been announced for the older model. The Next Web also says Nano Banana 2 is also called Gemini 3.1 Flash Image, in a February 26 preview sentence. The id gemini-3.1-flash-image and the alias “Gemini 3.1 Flash Image” look like the same object if you kebab-case the alias. Powerdrill printed an id. The Next Web printed a product name and, separately, an alias for the February preview. No shutdown date is the clause both pages share. A migration ticket that says “Nano Banana 2 shuts down on October 6” has invented a date neither page printed.

CNET’s October 12 pick is the practical consequence. The roundup was updated six days after the GA note, and the best-overall heading is still Nano Banana 2. A buyer who treats “best overall, Nano Banana 2” as a statement that 2.1 does not exist has ignored the October 6 notes. A buyer who treats the deprecation line as a statement that CNET retracted the pick has ignored the October 12 timestamp. Both pages can be current and still disagree about which name you should type.

$0.0336 is standard. $0.0168 is batch. $0.067 is the previous model.

Powerdrill’s table has two columns, Standard and Batch. Input, text, image, or video, per million tokens: $1.50 standard, $0.75 batch. Text and thinking output: $7.50 and $3.75. Image output: $30.00 and $15.00 per million tokens. One 1K image: $0.0336 and $0.0168. One 2K image: $0.0504 and $0.0252. One 4K image: $0.113 and $0.0567. The page says a 1K image consumes 1120 tokens, a 2K image 1680, and a 4K image 3780. At standard rates, 100 images at 2K come to about $5.04 before input and thinking tokens. 1680 times 100 is 168,000 tokens. At $30 per million image-output tokens, that is $5.04. The arithmetic matches the sentence. It does not include the input you sent, and it does not include thinking tokens. Powerdrill says thinking tokens are billed by default for thinking models. A budget that stops at $5.04 has used the image-output line and skipped the sentence under the table.

The Next Web does not list the batch column. It prints $0.0336 against $0.067 for Nano Banana 2, and $0.113 against $0.151. It also says input costs $1.50 per million tokens, up from $0.50, and that there is no free tier. “Up from $0.50” is a comparison to the older input price, not a third column. Powerdrill’s batch input is $0.75, not $0.50. A note that says “the cheap tier is $0.50” has merged The Next Web’s historical comparison with a batch price the same article did not print. Powerdrill is explicit that the free-tier column reads “Not available.” That matches The Next Web’s “no free tier” for the API. It does not match a Gemini app session, which is a different door. Powerdrill says the announcement does not list usage limits for each Gemini app plan, and that if you rely on the app you should check the limits in your own account.

Search grounding is a third meter, and only Powerdrill prints it in this set. 5,000 free search requests per month, shared across Gemini 3 and newer models, then $14 per 1,000 requests. That is not in The Next Web’s price paragraph. It is a meter you hit if you turn grounding on. The $0.0336 figure does not include it.

Half of $0.0336 is $0.0168. Half of $0.113 is $0.0565, and the table prints $0.0567. The page printed $0.0567. The 1K pair is a clean half. A router that assumes every batch cell is exactly half has checked one row.

Reference limits disagree. The surfaces disagree. The thinking dial is not the chat dial.

Powerdrill says Google’s pages give different limits for reference images. The API page says up to 14 reference images, with character consistency for up to 4 characters and object fidelity for up to 10 objects. DeepMind’s model page, in the Nano Banana family section, says up to five characters and up to fourteen objects. Five and fourteen are DeepMind’s. Four and ten are the API page’s.

The input list on the API page, as Powerdrill quotes it, is text, image, video, and PDF. The model returns images and text. Input token limit 131,072. Output token limit 32,768.

The rollout lists are not the same list. The Next Web says the model card names the Gemini app, Google AI Studio, the Gemini API, AI Mode in Search, Google Ads, Flow, and Stitch. Powerdrill says a follow-up X post named the Gemini app, AI Mode in Search, Google AI Studio, Flow, Stitch, Google Ads, and the Gemini Enterprise Platform. The API is on the model-card list and not, in that sentence, on the X-post list. Enterprise Platform is on the X-post list and not on the model-card list The Next Web used. If you need the API, cite the model card. If you need Enterprise, cite the post Powerdrill is quoting, and then check the tenant. A blog list is not an allowlist.

Thinking is two controls that share a shape and not a page. Powerdrill says the image model has configurable thinking levels, minimal, medium, or high, with medium as the default. The Next Web is talking about the Gemini app’s chat models when it describes low, medium, and high effort, and it quotes the help page: a higher level increases the model’s ability and uses more of your limit. Low is not minimal. The chat effort dial is not the image thinking dial. A config that sets “high” because both pages have a three-step control has assumed the enums match.

The preference scores are only in The Next Web. Nano Banana 2.1 with thinking scored 1050 for overall preference in text-to-image. Nano Banana 2 scored 990. Nano Banana Pro scored 935. Powerdrill’s DeepMind quotation is a qualitative line, that the update outperforms previous Nano Banana models on visual design, mask-based editing, subject consistency, and more natural-looking images. It does not print 1050. The known limits The Next Web lists are blurry small text at 1K and occasional left-right confusion. Those are limits on the model that just posted the higher preference score. A caption that says “1050, so the text is solved” has kept the score and dropped the 1K blur sentence. Powerdrill’s notes also say text rendering and infographic layout improved, and that tiling artifacts on 1:4, 4:1, 1:8, and 8:1 were fixed at 2K and 4K. Improved is not the same claim as “no blur at 1K.” The 1K blur limit and the wide-ratio fix can both be true. They are about different resolutions.

The free app door moved. The API never had a free column.

The Next Web’s lede is a chat-app change, not an image-price change, and the dates inside it do not agree with the day-of coverage it also cites. From the day of that article, free personal Gemini accounts move to Flash-Lite only, according to Google’s Gemini Apps Help page. The Next Web says this comes three days after the Nano Banana 2.1 release. October 6 plus three days is October 9, which is the date the search index gives the piece. AI Plus keeps Flash-Lite and Flash and loses Pro. AI Pro and AI Ultra keep all three. The same article says The Verge, on October 6, reported that free Gemini users could still choose Flash-Lite, Flash, and Pro. So the release-day report and the October 9 report describe different free-tier menus. October 6, three models still selectable, per The Verge as The Next Web cites it. October 9, free accounts moving to Flash-Lite only, per the help page as The Next Web cites it. A screenshot from the 6th does not refute a help-page change on the 9th. A help-page change on the 9th does not rewrite the 6th.

In the US, AI Plus is $4.99 a month, AI Pro is $19.99, and AI Ultra is $99.99 or $199.99. AI Pro subscribers will also get Deep Think, previously limited to Ultra, the site reported. 9to5Google’s October 3 report, as The Next Web summarizes it, named the models Gemini 3.5 Flash-Lite, Gemini 3.6 Flash, and Gemini 3.1 Pro. Those are chat model names. They are not gemini-nano-banana-2.1. The Next Web does not print that inclusion. It prints a chat-model cut, and a separate image-model release.

Earlier clocks in the same piece are context, not this week’s price. Nano Banana 2 previewed on February 26. On May 17, Gemini app users over 18 moved to compute-based limits that refresh every five hours until a weekly cap. Users under 18 moved on July 24. Media generation uses the limit faster. AI Plus gets twice the standard limits, AI Pro four times, AI Ultra five or twenty times the AI Pro limit depending on the plan. Those multipliers are about the app limit, not about the API’s $0.0336. A team that divides $0.0336 by four because they have AI Pro has applied an app multiplier to an API sticker. Powerdrill already said the announcement does not list per-plan app limits for this model.

The id to pin, if you are calling the API, is gemini-nano-banana-2.1. The price to pin, if you are not on batch, is $0.0336 for a 1K image and $0.113 for a 4K image, which is the pair both pages print. The name CNET’s October 12 heading still prints is Nano Banana 2. Three facts. None of them cancels the other.

What this means for you

If your image generation is not time-sensitive, batch mode halves the per-image cost. When migrating from Nano Banana 2, use the exact model ID from Google’s release notes and compare output quality on your own prompts before switching production traffic.