Qwen-Image 2.1 vs Nano Banana 2: What You Pay Per Image, and What You Pay to Own It

Only one of these two models has a price you can put in a spreadsheet. Nano Banana 2 — the Gemini 3.1 Flash Image model, generally available since 28 May 2026 — costs about $0.067 per 1024×1024 image, $0.101 at 2048×2048, and roughly $0.151 at the 4K end, with batch mode at about half and text input billed separately at $0.25 per million tokens. Qwen-Image 2.1, open-sourced by the Qwen-Image team on 20 September 2026, has no per-image price at all, because there is no hosted endpoint to buy from: it is a 33 GB weights download under a research licence that forbids commercial use. So the first question in this matchup is not which model is better. It is whether you are buying images or buying a machine that makes them, and those two purchases are not comparable in the way a spec sheet implies.

Forecastability is the thing a per-image price actually buys

Google's pricing is unusual in the image category for how cleanly it converts. The model bills $60 per million output tokens, and because image output is tokenised predictably, that resolves to per-image figures you can reason about before you generate anything.

• Nano Banana 2 — about $0.067 for 1024×1024, $0.101 for 2048×2048, roughly $0.151 at 4K, approximately half those figures in batch mode, plus $0.25 per million tokens for text input. A thousand 2K images is about $101, forecastable to the dollar.

• Qwen-Image 2.1 — no published price. Cost is your own GPU time against a 33 GB bundle, and the model card's advice for constrained hardware is CPU offload via pipe.enable_model_cpu_offload(), which is an escape hatch rather than a fix.

What the metered price buys is not just the images. It is the ability to answer "what does this feature cost us at ten thousand images a month" without benchmarking your own hardware first. That is a real advantage and it is easy to undervalue, because the alternative — self-hosting — has a cost that is genuinely hard to compute: it depends on your utilisation, your card, your electricity, and whether the box would otherwise be idle. The honest comparison is not $101 per thousand against free. It is $101 per thousand against an amortised figure you have to work out yourself, and most teams that do the arithmetic find the metered price is competitive right up until utilisation gets very high.

The licence is where "free weights" stops being free

This is the sharpest difference between the two models and it is not close.

Nano Banana 2 is a commercial product with commercial terms. You pay per image and you ship what you make. Qwen-Image 2.1 ships under the Qwen Research License Agreement dated 20 September 2026, which grants rights "FOR NON-COMMERCIAL PURPOSES ONLY" and states that commercial use requires a separate licence requested from the vendor. That is a change from the earlier Qwen-Image line, which shipped under Apache 2.0 — so the open option in this comparison is open in the sense that you can read it and run it, not in the sense that you can build a business on it.

For a research team, a hobbyist, or anyone benchmarking, that is a fine licence and the download is genuinely free. For a product team, it means the model in this matchup with no per-image price also has no commercial path, which collapses the cost comparison entirely: you are not choosing between $101 and something cheaper, you are choosing between a model you can sell output from and one you cannot.

Resolution and aspect ratios, where the open model has a real edge

Set the licence aside and Qwen-Image 2.1 does something Nano Banana 2 does not, in a way that matters for particular workflows.

• Qwen-Image 2.1 — native 2K with 2048×2048 as the documented default at 40 inference steps, seven aspect-ratio presets, up to 10 reference images, and RGBA output: a real alpha channel, transparent-layer editing, and subject extraction from RGB photos.

• Nano Banana 2 — 4K output with unusual ratio support, including 1:4 and 1:8 extremes, which is a longer reach than most models offer in either direction, and it publishes consistency figures of five characters and fourteen objects in a single generation.

The alpha channel is the one to notice. Google's model is not documented as producing transparency, and Qwen-Image 2.1 does so natively, which means the compositing step that would otherwise be manual work — cut out the subject, keep the soft edges, place it on something else — is inside the model rather than after it. If you build layered assets for a living, that is a workflow difference and not a quality claim. It is also worth saying plainly that both models render larger than most work needs: 4K and 2048×2048 are both past the point where the constraint is the model rather than the destination.

A published number against a promise

Google's model has been on both Artificial Analysis image boards since spring and sits inside the top five for text-to-image and image editing on the September snapshot, at around Elo 1,320 in text-to-image. Its consistency figures — five characters, fourteen objects — are Google's own, but they sit next to an independent scoreboard, which means they can be sanity-checked against how the model actually performs in blind comparisons.

Qwen-Image 2.1 has no independent score. It has a vendor blog post with a Qwen-Image-Bench chart that no third party has reproduced, and one hands-on early-access report from a tester in the Qwen Ambassador program who ran the final weights through a ModelScope Studio interface: roughly 10–15 seconds for text-to-image, 18–23 seconds for editing, strong camera-position and multi-character role handling, and multi-reference consistency degrading from about three input images onward. One reviewer, no timer shown, no published prompt set. It is the only outside observation of the model that exists in public, and it should be read as a hint rather than a measurement.

There is also a number circulating in third-party coverage describing Qwen-Image 2.1 as a 20-layer transformer. That contradicts the model card, which documents a 32-layer single-stream diffusion transformer with 7B parameters in the visual generation component, a Qwen3-VL 8B text encoder and a 64-channel RGBA VAE at 16× spatial compression. Use the model card.

The one thing you can do with both

This is the only article in this batch where the opponent is a model we actually route. Nano Banana 2 is on OrcaRouter as google/gemini-3.1-flash-image-preview, alongside the rest of the image line — OpenAI's GPT-Image family, Google's Imagen 4 tiers, the other Gemini image previews, and xAI's Grok Imagine image endpoint — all behind one OpenAI-compatible endpoint at provider list price with zero markup. Qwen-Image 2.1 is not among the models we route, and this piece is not going to imply that it is.

What that means in practice for this decision is narrower than a pitch and more useful than one. If you want to compare the two models on your own prompts, the Google side costs you an API key and about a tenth of a cent per image at 2K, and it is on the same endpoint as everything else you call. The Alibaba side costs you a 33 GB download, a GPU, and a research-licence review before you can do anything commercial with the output. Automatic failover across providers is the other piece that matters here: a route can carry production traffic on a measured model while an unmeasured one is evaluated beside it, so the evaluation never sits on the critical path.

Which one, for whom

• Choose Nano Banana 2 if you are shipping. It has a commercial licence, a price you can forecast to the dollar, a top-five independent score on both boards, 4K output, and the widest ratio range in this comparison. The cost of a thousand 2K images is about $101 and you can be generating them this afternoon.

• Choose Qwen-Image 2.1 if you are researching. It is free to download, the architecture and the prompt-rewriting system prompt are fully inspectable, it renders natively at 2048×2048 with real transparency, and it takes up to 10 reference images with local edits by circle, painted annotation or mask. You cannot sell what it produces, and nobody has measured whether it is any good.

The gap between those two paragraphs is the gap between the models, and it is not a quality gap — it is a maturity gap, and it is closing on a schedule. Qwen's early-access testers were asked to publish samples or reviews by 29 September 2026. When those land, the open model stops being an unknown and starts being comparable, and the only question left will be the licence. That is the number to watch, and it is a licence file rather than a benchmark.