Nano Banana 2.1 API
AI ImageThe Nano Banana 2.1 API generates and edits images on OmniaKey, with a per-call price for each output size.
- Per successful call$0.019–$0.039
- OperationsText to image
- Image edit
- View tool pricing and billing rules
Your result will appear here once you run the model.
01
Where it fits
The Nano Banana 2.1 API generates and edits images on OmniaKey, with a per-call price for each output size.
Choose 1K, 2K or 4K to match your deliverable, then compare results on your own prompts.
02
What it does
Text to image
gemini-nano-banana-2.1
Turn a text prompt into a new image.
Image edit
gemini-nano-banana-2.1
Change an existing image from a prompt and reference input.
03
In depth
Nano Banana 2.1 is Google's Gemini image model on OmniaKey, sold as gemini-nano-banana-2.1 with text-to-image and image-edit capabilities. The catalog prices it per call by output size across 1K, 2K and 4K, the same listed tier as Nano Banana 2 and a lower one than Nano Banana Pro at every size. Google presents it as the successor to Nano Banana 2, pointing at sharper typography, steadier multi-turn edits and a fix for panoramic tiling, and recommends it for new projects.
It is a reasoning model, which shapes how you plan a queue. Each of our five generations spent between 812 and 1,011 thinking tokens, 860 at the median, and thinking defaults to medium here while Nano Banana 2 defaults to minimal. The output justified some of that: in our matched runs it won a Chinese tea-house sign and a fish-market portrait, took an infographic and a product shot on the only clean pair of runs, and tied a product edit.
Two kinds of failure turn up in our logs. Punctuation gets copied rather than interpreted, so the em dashes we drafted survived into a cafe menu and pushed the price column out of line with the dish names. Detail gets quietly dropped: in one new-scene test it kept a person's look but hid the scar the reference showed, and in a poster run it added a tagline nobody asked for. Check the words and the small marks before you print.
Candid people and places
Give the frame a location and a moment instead of a pose list. Our fish-market portrait won on the details around the subject: Lisbon tiles, a Portuguese price card and a background blurred the way a 50mm lens would blur it. Name the lens, the light and the skin texture, and let the setting carry the story.
Infographics with numbers
Charts and labelled diagrams are a strong fit: this was the only model in our six-test round with both infographic runs clean, while a sibling misspelled a label. Keep the data plain in the prompt, give every bar and number a fixed position, and verify each figure by eye at delivery size.
Three-round edits
Change a colour, then a backdrop, then a headline, one task at a time. Google points at conversational consistency for exactly this pattern, and our product edit came out as a tie with the object intact. Feed the previous result in as the next source and record the thinking level beside each attempt.
04
Hands-on notes
In our four-prompt test against Qwen Image 3, Nano Banana 2.1 won the Chinese tea-house sign and the Lisbon portrait and tied the product edit.
Read the testIn the same test Nano Banana 2.1 was slower on new images than Qwen Image 3, at a median 36.5 seconds against 23.2.
Read the testOur five Nano Banana 2.1 generations spent between 812 and 1,011 thinking tokens each, with a median of 860 per image.
Read the testOn our cafe-menu prompt, Nano Banana 2.1 kept the em dashes we had drafted, so the price column never lined up with the item names.
Read the testIn our living-room edit, Nano Banana 2.1 kept the bookshelf but shortened the sofa to fit the new plant into the frame.
Read the test
05
Limits to plan around
- A flat poster can pick up a line you did not write: our run returned the sheet as a flat file, as asked, but appended a tagline, so scan for extra copy before printing.
- Punctuation is reproduced literally. The em dashes in our menu draft came back intact and pulled the price column away from the dish names, so write plain separators.
- Faces lose small marks. In our new-scene test the person's look survived but the scar did not, so compare faces at full size whatever model made them.
- Thinking stays on at medium unless you change it: our five runs spent 812 to 1,011 thinking tokens each, which is reasoning work a simple brief does not need.
06
Compare nearby models
| Model | Best for | Operations | Listed price |
|---|---|---|---|
| Nano Banana 2.1This page gemini-nano-banana-2.1 | New Gemini image work where lettering, place cues and a 4K option all matter at once. | Text to image · Image edit | $0.019–$0.039 per call |
| Nano Banana Pro gemini-3-pro-image | Complex compositions and reasoning-heavy edits, one listed tier higher at every size. | Text to image · Image edit | $0.039–$0.069 per call |
| Nano Banana 2 gemini-3.1-flash-image | The previous generation at the same listed tier, deprecated by Google but still serving requests. | Text to image · Image edit | $0.019–$0.039 per call |
07
Sources and verification
Guide updated 2026-10-27.
08
Call it
Media models run asynchronously. Create a task, poll it until it finishes, then download the signed asset URL.
# 1. Create the job. 202 Accepted; the body carries the task id,
# the pre-charged price, and a poll_url. Reuse one Idempotency-Key
# across retries so a timeout never starts a second paid job.
curl --fail-with-body https://api.omniakey.com/v1/media/tasks \
-H "Authorization: Bearer $OMNIAKEY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: gemini-nano-banana-2.1-001" \
-d '{
"model": "gemini-nano-banana-2.1",
"type": "image",
"operation": "generate",
"input": {"prompt": "a paper boat on a quiet lake at sunset"},
"parameters": {"aspect_ratio": "1:1"}
}'
# 2. Poll GET /v1/media/tasks/{id} until status is
# succeeded | failed | canceled. Progress carries percent and stage.
curl --fail-with-body https://api.omniakey.com/v1/media/tasks/TASK_ID \
-H "Authorization: Bearer $OMNIAKEY_API_KEY"
# 3. Download assets[0].url before expires_at, or poll again for a
# fresh signed URL. Failed and canceled jobs are refunded.
curl -sS -o out.png "$(jq -r '.assets[0].url' result.json)"09
Pricing
Image jobs are billed per successful call, not per token. Failed or canceled jobs are refunded automatically.
$0.019–$0.039per callSave 43%
image_size
- 1K
- $0.019per call
- 2K
- $0.029per call
- 4K
- $0.039per call
More Google image models
Related articles
GPT Image 2.6 vs Nano Banana 2.1: Image Quality and Cost
Four tasks, 15 images: what Sunburst, Flare and Nano Banana 2.1 returned, and what each costs.
Read articleNano Banana Product Photography: 8 Prompts Tested
Eight prompts, 22 edits, two models: which shots kept the label, where the shoe got mirrored and which model to pick for each job.
Read articleGPT Image 2.5 vs Nano Banana 2.1: Same-Prompt Test and Price
Four prompts, three image models, one day: what each made and what it really cost.
Read articleQwen Image 3 vs Nano Banana 2.1: Same-Prompt Test and Price
Four prompts, two image models, one day: what each made and what it cost.
Read articleFrequently asked questions
How do I call Nano Banana 2.1 on OmniaKey?
POST to /v1/media/tasks with the model id, type "image", and operation "generate" or "edit". Reuse one Idempotency-Key across retries so a timeout never starts a second paid job.
What does Nano Banana 2.1 cost?
Nano Banana 2.1 bills $0.019–$0.039 per call. Creation pre-charges that amount from your prepaid balance, and failed or canceled jobs are refunded automatically.
Why does the API return a task id instead of the image?
Provider latency runs from seconds to about a minute, so OmniaKey normalizes every image route into one asynchronous lifecycle. Poll GET /v1/media/tasks/{id} until the status is succeeded, then read the signed asset URL.
Nano Banana 2.1 or Nano Banana Pro?
Pro sits one listed tier higher at every output size and is built for complex layouts and reasoning-heavy edits, while 2.1 is the general-purpose choice. In our six-test round 2.1 won the infographic and the product shot, Pro won the photoreal scene, and Pro tied Nano Banana 2 on keeping a person's details.
What does the model id look like in a request?
It is gemini-nano-banana-2.1, case-sensitive, with generate and edit as the two operations. It is not an alias for gemini-3.1-flash-image: Google released it separately on 6 October 2026. Its model page also lists video and PDF as inputs, though the response is still an image plus text.
Which output sizes can I ask for?
1K, 2K and 4K, with 1K as the default. The 512 setting belongs to the previous Nano Banana 2 and is not on this model's page, so check that during a migration. A 4K label names an output tier rather than a square canvas, so state the aspect ratio you want as well.
How many reference images does it accept?
Google lists up to 14, split into 10 objects and 4 characters, and those are separate ceilings rather than a promise that 14 people stay recognisable. Send references in order of importance, and check faces, logos and packaging after each round of edits.
Does every image spend thinking tokens?
Yes. Thinking runs at medium unless you change it, and our five generations spent 812 to 1,011 thinking tokens each, 860 at the median. Set the level explicitly when you compare models, because Nano Banana 2 defaults to minimal, and run both ids at one level before reading anything into the difference.
