Qwen Image 3 vs Nano Banana 2.1: Same-Prompt Test and Price

We ran the same four prompts through Qwen Image 3 and Nano Banana 2.1 on October 25, 2026, and both spelled every English and Chinese word right. Nano Banana 2.1 won two tests, Qwen Image 3 won one, and one was a tie. A 2K image costs $0.03 at Alibaba's list price and about $0.057 at Google's, thinking included.

Still on Nano Banana 2? Google deprecated it on October 6, 2026, and 2.1 costs half as much per 2K image.

Both are on OmniaKey: Qwen Image 3 and Nano Banana 2.1.

How we tested

• Models and IDs: qwen-image-3 and gemini-nano-banana-2.1 on OmniaKey. OmniaKey's qwen-image-3 is Alibaba's qwen-image-3.0, so Alibaba's docs and prices below name that ID.
• Parameters we sent: Qwen Image 3 got {"size": "2048*2048", "prompt_extend": false, "watermark": false} and Nano Banana 2.1 got {"aspect_ratio": "1:1", "image_size": "2K"}. Every image is 2048×2048.
• All runs on October 25, 2026. The poster and the Chinese sign ran twice per model; the portrait and the edit ran once. That is 12 test images plus one edit source, and every task succeeded on the first try.
• Fairness: Qwen's prompt rewriting (prompt_extend, on by default) was off, which also turns off its thinking mode, so Qwen drew our exact wording. Nano Banana 2.1 kept its default thinking.
• The edit source photo was made with GPT Image 2 at 1024×1024, so neither model edited its own image.
• Time is from sending the job to the task showing succeeded.
• We judged at full size: every requested word letter by letter, then whether the image follows the prompt, meaning layout, place and depth of field, and for the edit whether the product and its label survived.
• OmniaKey sells API access to both models, so we ran the tests through it. Every price below is Alibaba's or Google's own list price, not ours.
• Sources: Alibaba's Qwen image API reference, Model Studio pricing and newly released models; Google's Gemini API pricing, image generation and changelog, each checked October 25, 2026.

One or two images per prompt is a small sample, so read the results as examples rather than a benchmark.

Four tests, same prompts

Test 1: English jazz poster (1:1)

This test checks every word of a mixed-case poster line, and whether the file is a poster or a photo of one.

Prompt:

A minimalist concert poster for a jazz night. Large headline at the top: BLUE HOUR SESSIONS. Below it, in smaller type: Friday 14 November · 8 PM · The Lantern Room. In the middle, a brass saxophone silhouette against a deep blue gradient. Off-white paper texture, generous margins, clean modern typography.

• Qwen Image 3 (23.1 s and 23.5 s): every word right in both runs. The small line reads "Friday 14 November · 8 PM · The Lantern Room" in the same mixed case as the prompt. Off-white paper, a flat full-frame poster, and a near-black saxophone silhouette on a blue gradient panel both times.
• Nano Banana 2.1 (32.2 s and 36.5 s): every word right in both runs, but the small line is set in capitals: "FRIDAY 14 NOVEMBER · 8 PM · THE LANTERN ROOM". The saxophone is gold and brass-colored, on a blue circle in run 1 and a blue panel in run 2. Both runs show the poster as a sheet of paper photographed on a grey or dark background, so you have to crop it before use.

Verdict: Qwen Image 3. Both spelled every word right, twice. Qwen kept our capitalization and gave a flat poster ready to print. Nano's brass saxophone is closer to the prompt, but it changed the small line to capitals and framed the poster as a photographed mockup both times. For posters on Nano, add "flat design, no mockup, keep the capitalization".

Test 2: Chinese tea-house sign (1:1)

This test checks two lines of Chinese on a signboard, and whether the result looks like the realistic photo the prompt asks for.

Prompt:

一家老茶馆的门面，傍晚暖光，写实摄影。门楣上挂着一块深色木质牌匾，金色大字写着“清风茶舍”，牌匾下方一行小字“始于一九八六”。门两侧各挂一盏红灯笼，门口摆着几把竹椅。

In English: the front of an old tea house, warm evening light, realistic photography. A dark wooden plaque over the door with large gold characters reading 清风茶舍 (Qingfeng Tea House) and a smaller line below, 始于一九八六 (since 1986). A red lantern on each side of the door, a few bamboo chairs outside.

• Qwen Image 3 (23.1 s and 23.2 s): "清风茶舍" and "始于一九八六" are both correct in both runs. Glossy gold characters on clean dark wood, roof tiles above, warm light, red lanterns, bamboo chairs. The finish reads more like a 3D render than a photo.
• Nano Banana 2.1 (36.5 s and 32.0 s): both lines correct in both runs. The plaque is weathered wood with peeling paint and matte gold characters. It reads like a real photo: people inside, lit lanterns, plants, a tea set, even an air conditioner unit on the wall.

Verdict: Nano Banana 2.1. All four images have both Chinese lines right, so text is a tie. The prompt asked for 写实摄影, realistic photography, and Nano's version looks like a photo while Qwen's looks like a clean render.

Test 3: Lisbon fish-market portrait (1:1)

This test checks a candid portrait: place cues, shallow depth of field and skin texture.

Prompt:

Candid street portrait of a woman in her sixties laughing at a fish market stall in Lisbon, holding a paper-wrapped bundle of sardines in both hands. Morning light, shallow depth of field, natural skin texture with fine wrinkles, ice and silver fish blurred in the background. Shot on a 50mm lens.

• Qwen Image 3 (27.4 s): a laughing woman with auburn hair holding paper-wrapped sardines in both hands. An indoor fish counter with ice and silver fish behind her, lightly blurred. Nothing in the frame says Lisbon.
• Nano Banana 2.1 (37.1 s): a laughing grey-haired woman in a scarf holding a paper bundle of sardines in both hands. Blue-and-white tiles on the wall, a handwritten price card reading "SARDINHAS 1,40", people blurred in the background, a stone floor.

Verdict: Nano Banana 2.1. It put the scene in Lisbon, with tiles and a Portuguese price card, and blurred the background more, which is what "shallow depth of field" asks for. Qwen's portrait could come from any indoor fish market.

Test 4: Edit a product photo (1:1)

This test checks whether an edit changes the background only, with the bottle and its label left exactly as they were. The source photo was made with GPT Image 2 and both models got the same file.

Prompt:

Place this exact bottle on a sunlit white marble kitchen counter with a blurred window behind it and a sprig of rosemary beside it. Keep the bottle, the cork, the label and every word on the label exactly the same.

• Qwen Image 3 (41.5 s): kept the bottle shape, the cork and the label, "OLIVA", the olive sprig drawing and "Extra Virgin · 500 ml". Marble counter and rosemary beside the bottle. The window behind is sharp. The bottle looks closest to the source.
• Nano Banana 2.1 (28.4 s): kept the cork and the same label words and drawing. Marble counter, rosemary, and a strongly blurred window behind, as the prompt asked. The oil looks a little more golden than in the source, and the whole frame looks more like a photograph.

Verdict: Tie. Both kept every word on the label. Qwen kept the bottle closest to the source. Nano followed "blurred window", looks more photographed, and was faster on this edit. Pick Qwen when the product must not change at all.

Results by test

Text did not decide: all eight poster and sign generations spelled every requested word right, English and Chinese. Case, layout and how photographic the result looks did. Qwen followed the wording literally. Nano added place detail such as tiles and a price card.

Price and speed

Alibaba's list price, per Alibaba Cloud Model Studio's pricing page, checked October 25, 2026:

Google's list price, per the Gemini API pricing page, checked October 25, 2026:

What one image from our runs costs at those list prices:

Qwen's edit is $0.03 plus $0.003 for the input image, with no token charge. Nano Banana 2.1's five generations used 812, 825, 860, 878 and 1,011 thinking tokens, a median of 860, which adds 860 × $7.50 per million = $0.0065 to the $0.0504 image. Our Nano edit used 87 thinking tokens and a 1,120-token input image.

The gap comes from thinking. Google bills thinking tokens as text output at $7.50 per million and leaves thinking on at medium by default, so the per-image figure on its price table is a floor rather than the bill.

Qwen was faster on every new image, 23.1 to 27.4 s against 32.0 to 37.1 s for Nano, and slower on the one edit. For OmniaKey's current rates, see the Qwen Image 3 and Nano Banana 2.1 model pages.

What about Nano Banana 2?

Google deprecated Nano Banana 2 (gemini-3.1-flash-image) on October 6, 2026, the day it released 2.1, per its changelog checked October 25, 2026. It names 2.1 as the replacement and has announced no shutdown date. Google's docs still list Nano Banana 2 as stable and supported, though they point new projects to 2.1.

On price alone, 2.1 is the move: $0.0504 per 2K image against $0.101 for Nano Banana 2, and $0.0252 against $0.050 on the Batch API. If you searched "qwen image 3 vs nano banana 2", test 2.1: Google now points new projects to it, and the sign and portrait results above are 2.1's. The differences between 2 and 2.1, and how 2.1 lines up with Pro, are in our Nano Banana 2.1 vs Pro test.

Pros and cons

Qwen Image 3

• Pros: exact text including capitalization; flat layouts with no mockup; fastest generations at a 23 s median; lowest list price per 2K image at $0.03.
• Cons: clean, render-like look in photo prompts; few place details unless you ask for them; slowest on the edit at 41.5 s; tops out at 2048×2048.

Nano Banana 2.1

• Pros: most photographic results; adds place and real-world detail; faster edit; up to 4K; Batch API at half price.
• Cons: changed the case of our small line; framed the poster as a mockup both times; slower generations at a 36.5 s median; spends thinking tokens on every image.

Which one for which job

Run the same test through one API

This request sends the poster prompt to Qwen Image 3 at 2048×2048 with prompt rewriting and the watermark off, as we ran it:

curl https://api.omniakey.com/v1/media/tasks \
  -H "Authorization: Bearer $OMNIAKEY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: blue-hour-poster-001" \
  -d '{
    "model": "qwen-image-3",
    "type": "image",
    "operation": "generate",
    "input": {"prompt": "A minimalist concert poster for a jazz night. Large headline at the top: BLUE HOUR SESSIONS. Below it, in smaller type: Friday 14 November · 8 PM · The Lantern Room. In the middle, a brass saxophone silhouette against a deep blue gradient. Off-white paper texture, generous margins, clean modern typography."},
    "parameters": {"size": "2048*2048", "prompt_extend": false, "watermark": false}
  }'

1. The reply is 202 with a task id.
2. Every few seconds, fetch https://api.omniakey.com/v1/media/tasks/{id} until status is succeeded.
3. Download the image from assets[0].url.

For Nano Banana 2.1, set model to gemini-nano-banana-2.1 and parameters to {"aspect_ratio": "1:1", "image_size": "2K"}. To edit, set operation to edit: Qwen takes the source in input.images as an array of URLs or data URLs, and Nano Banana 2.1 takes input.image as {"data": "<base64>", "mime_type": "image/jpeg"}.

Open the Qwen Image 3 and Nano Banana 2.1 model pages to start a run, and the image API docs for every parameter.

Common questions

Is Qwen Image 3 better than Nano Banana 2.1?

For posters and other flat layouts where the exact text matters, yes: it won the poster test with a flat file in 23 s. For photos, Nano Banana 2.1 is better: it won the Chinese storefront and the Lisbon portrait.

Is Qwen Image 3 open source?

No. There are no Qwen Image 3.0 weights: the Qwen team's Hugging Face page lists the Qwen-Image-2.1 family and nothing for 3.0. The newest open Qwen image model is Qwen-Image-2.1, 7B, under the Qwen Research License. Our Qwen-Image-2.1 review covers what those weights do on your own machine.

Can Qwen Image 3 make 4K images?

No. Its maximum output is 2048×2048, with aspect ratios from 1:8 to 8:1, so 2K is the ceiling. Nano Banana 2.1 outputs 1K, 2K or 4K, and Google charges $0.113 for a 4K image.

Which is cheaper, Qwen Image 3 or Nano Banana 2.1?

At the vendors' list prices, Qwen Image 3 is about half: $0.03 for a 2K image against $0.0504 for Nano Banana 2.1, or about $0.057 with its thinking tokens counted. Edits come to $0.033 on Qwen and about $0.053 on Nano.

Does Nano Banana 2.1 add a watermark?

Every Nano Banana 2.1 image carries a SynthID watermark, which is invisible, per Google's image generation docs. Qwen Image 3 adds a watermark only if you ask for one, because watermark defaults to false.

Update log

• October 25, 2026: first version, four same-prompt tests run the same day.