Gemini Nano Banana 2.1 vs Qwen Image 3.0: A Tie at 1K, a Split at 2K

Put Gem​ini Nano Banana 2.1 and Qw​en Image 3.0 on a rate card and the first number is funny: a 1K image from the newest Gem​ini image model costs $0.0336, and a 1K generation from Qw​en Image 3.0 on Model Studio costs $0.03438. Seven hundredths of a cent apart. Two labs on opposite sides of the world, publishing in different currencies for different platforms, landed on the same price for the thing most people actually buy. The tie breaks the moment you move to 2K — $0.0504 against $0.0688, a 36% gap in Gem​ini's favour — and it breaks differently again once you notice that Qw​en bills image input and editing input as separate meters while Nano Banana 2.1 does not. Which of those two facts decides your build depends entirely on how many times your pipeline touches an image before it ships one.

They are also more different than the price suggests. Nano Banana 2.1 went generally available on October 6, 2026 as a hosted API model with a 23-day migration deadline attached to its predecessor. Qwen Image 3.0 launched on Alibaba's Qwen AI platform on 5 August 2026 with Standard and Pro tiers and APIs at the same time, and reaches international developers through Qwen Cloud and Alibaba Cloud Model Studio. One is a Western frontier-lab model priced per image token; the other is a Chinese platform model priced per image, with a published concurrency budget and a domestic and international rate card that do not agree with each other.

The two rate cards, line by line

Nano Banana 2.1 is billed in tokens with a per-image conversion published alongside. Output is $30 per million image tokens, which Google resolves to $0.0336, $0.0504 and $0.0756 for 1K, 2K and 4K respectively, with the per-image token counts given as 1,120, 1,680 and 2,520. Input is $1.50 per million tokens across text, image and video. Batch output is half: $0.0168, $0.0252 and $0.0378. There is no separate meter for feeding an image back in — an image you pass as input is counted at the input token rate, once.

Qwen Image 3.0 comes in two tiers on a Chinese cloud rate card. Standard is RMB 0.18 per image at both 1K and 2K, with image input at RMB 0.02 and editing input at RMB 0.02 per image. Pro is RMB 0.25 at 1K and RMB 0.50 at 2K. On Qwen Cloud international the Standard tier runs $0.003 input and $0.03 output per image. On Alibaba Cloud Model Studio in USD the published figures are $0.00275 per 1K input, $0.03438 per 1K generation, $0.068761 per 2K generation, $0.04 per 1K output and $0.075 per 2K output. Throughput is published too: 10 concurrent requests, a 200-request asynchronous queue and 20 requests per minute.

• 1K generation — Gemini Nano Banana 2.1 $0.0336; Qwen Image 3.0 Standard $0.03438 on Model Studio. Effectively a tie.

• 2K generation — Nano Banana 2.1 $0.0504; Qwen Image 3.0 Standard $0.068761. Google is 36% cheaper.

• 4K — Nano Banana 2.1 $0.0756 and it is a documented tier; Qwen Image 3.0's Model Studio entry caps total pixels at 2048×2048, so there is no 4K row to compare.

• Image input — Nano Banana 2.1 counts it at the $1.50 per million input rate; Qwen Image 3.0 meters it separately at $0.00275 per 1K input, with editing input charged at RMB 0.02 domestically.

• Batch — Nano Banana 2.1 halves everything through the Batch API at up to 24-hour turnaround; Qwen Image 3.0's published lever is an asynchronous queue of 200 rather than a discount tier.

• Throughput — Qwen Image 3.0 publishes 10 concurrent, 200 queued, 20 RPM; Google publishes none of these for Nano Banana 2.1.

• Domestically — Qwen Image 3.0 Standard is RMB 0.18 per image on the Chinese rate card; Nano Banana 2.1 has no domestic Chinese tier because Google does not sell one.

Two of those rows matter more than the rest. The 4K row is a capability gap rather than a price gap — if you need 4K out, Qwen Image 3.0's Model Studio entry is not in the conversation at all. And the image-input row is where the two billing philosophies diverge sharply, which is worth its own section because it is the part that does not show up in a headline price.

Why the second pass changes the maths

A single-shot generation is the easy case, and on a single shot these two are interchangeable on price. Real image work is rarely single-shot. A typical editing workflow generates a candidate, feeds a reference image back in for a correction, feeds the corrected image back again for a style pass, and only then ships. Every one of those returns is an image input.

On Nano Banana 2.1 that input is metered at the general input rate alongside the text prompt, and the model's conversational editing is explicitly designed for it — multi-turn editing via a previous interaction is a documented feature, and thinking mode carries the conversation. On Qwen Image 3.0 the same loop is metered on a dedicated line the vendor publishes: image input at $0.00275 per 1K on Model Studio, RMB 0.02 per image as editing input domestically. Neither is expensive per call. The difference is that Qwen's meter makes iteration cost visible as its own line item while Google's folds it into input tokens, and a team budgeting from the headline per-image figure will underestimate the three-pass workflow on both platforms.

Run the arithmetic honestly and the practical guidance is boring but real. At the 1K price they are the same, so the price is not the reason to pick either. At 2K, Nano Banana 2.1 is meaningfully cheaper per generation and additionally offers a batch tier that halves it again. If you need 4K, there is no comparison to run. If you need a published concurrency budget you can plan a queue around, Qwen Image 3.0 publishes one and Nano Banana 2.1 does not. That is the shape of the decision: mostly about the resolution and the throughput guarantee you need, not about the price at the tier everyone quotes.

What is measured, and for whom

The evidence asymmetry here runs opposite to what you might expect. Qwen Image 3.0 has an independent placement and Nano Banana 2.1, at a day old, does not. On Artificial Analysis's text-to-image board, from samples collected in July 2026, Qwen-Image-3.0 sits at Elo 1,077 in sixteenth place from 5,816 samples and Qwen-Image-3.0-Pro at 1,088 in fourteenth from 6,077 — mid-table, well-sampled, and about where a strong platform model lands rather than a frontier one. Alibaba's own launch claim, that the model took the number one spot in China on Arena.ai's text-to-image leaderboard, is a vendor-reported figure on a different board with a different voter population; it does not contradict the Artificial Analysis placement, it just measures something else.

Nano Banana 2.1's side of the ledger is thinner. There is no Elo, no vote count and no third-party render comparison in the public record as of October 7, 2026. What exists is Google's changelog language — "significant improvements" in visual quality, prompt adherence, multi-turn character consistency and text rendering — and the documented feature changes: 1K, 2K and 4K output with no 512px tier, up to 14 reference images split as ten objects and four characters, thinking mode on by default, and Google Image Search grounding alongside web search. If you want to compare these two on quality today, you are comparing one model's measured mid-table score against another model's vendor claim, and those are not the same kind of number.

Where you can call each of them

Neither model is on OrcaRouter, and it is worth saying so plainly rather than implying otherwise. Gemini Nano Banana 2.1 is available through the vendor's own API and from several third-party platforms; Qwen Image 3.0 is available through its own platform and several third-party platforms. Our catalogue does carry other image models — OpenAI's openai/gpt-image-2 and google/gemini-3.1-flash-image-preview among them — and 207 models in total behind one OpenAI-compatible endpoint.

That matters to this comparison for one specific reason: both of these vendors change their numbers. Google moved Nano Banana 2.1's predecessor's whole tier structure and announced a shutdown on the same day; Alibaba publishes one rate card in RMB for the domestic platform, a second in USD for Model Studio and a third set of figures for international Qwen Cloud. When a vendor's price is a moving target, the value of buying through a layer that passes provider list price through at 0% markup is that the number on your side moves the same day it moves on theirs, with no margin between you and the published figure and no separate contract to renegotiate. That is the discipline OrcaRouter's routing is built on, and it is the reason a rate-card comparison like this one stays true a month later instead of drifting.

Which one to actually pick

Choose Gemini Nano Banana 2.1 when you need 4K output, when you want a batch tier that halves the per-image cost for large reference passes, when prompt-based editing with up to 14 reference images is the workflow, or when you want Google Search grounding inside the generation. The migration deadline on its predecessor is a reason to move to it rather than a reason to avoid it, and the price drop from Nano Banana 2 is roughly a halving at every resolution. The open questions are the absence of any independent quality measurement and the unpublished throughput limits, which is why a production path should be prepared to fail over rather than assume the endpoint holds.

Choose Qwen Image 3.0 when the output stays at or below 2048×2048, when a published concurrency budget matters because you are queueing work, when the Chinese rate card at RMB 0.18 per image is the commercially relevant number, or when the Pro tier's higher price buys you something your evaluation can demonstrate. The mid-table independent placement is not a weakness for this use — it is the most reliable single fact in this comparison, because it is the only number here that neither vendor produced.

If you have the luxury of a bake-off, run it on the dimension that actually differs: the multi-pass edit. Generate a candidate, then do three successive corrections on each model, and count the total cost of the accepted image rather than the cost of the first render. That is the comparison neither vendor's rate card performs for you, and it is the one that decides whether a "tie" at 1K is really a tie.