Qwen-Image-2.1 vs GPT Image 2: Download It Free, or Rent the Best One

Qwen-Image-2.1 is a 33 GB download you can run for nothing and cannot sell. GPT Image 2 is an API you cannot download at all, billed by the token, four months into production and sitting at the top of the independent image leaderboards. Those two facts are the entire trade, and the interesting part is that the transparency feature — the thing that makes Qwen-Image-2.1 worth a look in the first place — arrived on both systems within a month of each other by completely different routes. Qwen-Image-2.1 shipped on 20 September 2026 with alpha built into its latent space. GPT Image 2 gained a background: "transparent" parameter on 20 August 2026 as an API flag.

Two routes to the same feature

Transparent output is the reason most people end up comparing these two, so start there, because the implementations are not equivalent and the difference matters more than the output.

Qwen-Image-2.1's transparency is architectural. A 64-channel RGBA VAE with 16× spatial compression means alpha is part of the latent space the model denoises in. The same mechanism gives you generation with transparency, editing inside an image that already has transparency without flattening it, and subject extraction from an ordinary RGB photograph that returns alpha rather than a binary mask. It is one capability with three uses, and the vendor's own framing is that no separate background-removal node, matting model or edge-cleanup pass is needed.

GPT Image 2's transparency is a request parameter. Adding background: "transparent" alongside output_format: "png" returns an image with a real alpha channel, and it works across text-to-image, image editing, and the image generation tool in the Responses API. Mask-based inpainting and transparent backgrounds can be combined. It shipped in Preview, so OpenAI's own guidance is that results may be unstable — and editing with transparency is described as redrawing or removing a background rather than precise outline matting. At least two secondary sources claim the parameter is unavailable and that GPT Image 2 does not output transparency at all; those contradict the announcement coverage, the API documentation mirrors, and third-party testing, so treat them as stale or wrong rather than as a counter-signal. Transparency does not change the token cost — at a given quality and size, a transparent image and an opaque one consume the same image tokens.

So both produce alpha. One does it because the model was built that way; the other because a parameter asks the service for it. Which is better depends entirely on whether the alpha has to survive an edit loop, and that is testable in an afternoon.

What GPT Image 2's head start buys you

Four months in production is not a marketing point, it is a maturity gap, and it shows up in places that are boring and decisive.

• Independent standing — GPT Image 2 sits at the top of the independent image leaderboards and has been there long enough for that to be a stable claim rather than a launch-week spike. Qwen-Image-2.1 was absent from those boards entirely when this was written.
• Documented failure modes — a model with four months of public use has a body of known failure cases you can read before you hit them. A model that went public yesterday has a vendor scorecard and one integration test.
• Operational surface — tiered rate limits expressed in both tokens per minute and images per minute (from 100,000 TPM and 5 IPM at the entry tier up to 8,000,000 TPM and 250 IPM at the top), plus batch endpoint support. That is the difference between a demo and a queue you can size.
• Quality figures from a third party — the leaderboard position is independently measured. Qwen-Image-2.1's only number is the vendor's own Qwen-Image-Bench chart, where it reads 60.28 against Nano Banana 2.0 at 59.82 and GPT Image 1.5 at 59.65. Vendor material, unreproduced.

The one genuinely independent data point on Qwen-Image-2.1 is the functional verification published with the SGLang integration: transparent PNG output passed, alpha was retained across the full 0–255 range, RGBA PSNR measured 60.69 dB on a single sample, and FP8 configurations passed transparent-PNG generation. The SGLang authors say explicitly that this is a correctness check and not a general quality guarantee. It is the strongest evidence available that the alpha channel is real. It says nothing about whether the pictures are good.

The bill, worked through

GPT Image 2 is token-billed, which means the cost of an asset depends on quality tier and resolution in a way that a per-image price hides. Published list rates are $5.00 per million text input tokens, $8.00 per million image input tokens, and $30.00 per million image output tokens, with cached rates at $1.25 and $2.00 respectively. OpenAI does not publish a static output-token table for this model — the older table covering Low, Medium and High token counts is explicitly marked as not applying to GPT Image 2 — so the numbers below are third-party measurements of list pricing at 1:1 aspect ratio, text-to-image:

• 1K output — low quality $0.0059, medium $0.053, high $0.211 per image.
• 2K output — $0.012 low, $0.107 medium, $0.428 high.
• 4K output — $0.020 low, $0.178 medium, $0.712 high.

The spread between low and high quality at the same resolution is a factor of roughly thirty-five. That is the actual decision inside a GPT Image 2 pipeline: not which model, but which quality tier survives review at the lowest cost. And it interacts with editing in a way that catches people out — input_fidelity cannot be adjusted on this model because it processes every image input at high fidelity automatically, so an edit request carrying reference images consumes more image-input tokens than you would estimate from the output size. Generate-inspect-edit loops are therefore more expensive here than the per-image figures suggest.

Against that, Qwen-Image-2.1's marginal cost per image is electricity. The catch is the fixed cost and the licence. Thirty-three gigabytes to download, a DiT at roughly 14 GB with the rest going to the Qwen3-VL 8B text encoder, CPU offload recommended on constrained cards, and the whole thing under the Qwen Research License Agreement dated 20 September 2026 — non-commercial only, with commercial rights by separate arrangement and no published price or terms for them.

Where the hosted option sits

There is a middle path that avoids both the GPU and the licence question, and it is the one most teams actually take. OrcaRouter routes OpenAI's GPT-Image family alongside Google's Imagen 4 tiers and Gemini image previews and xAI's Grok Imagine image endpoint — 200+ models behind one OpenAI-compatible endpoint, provider list price passed through at zero markup, automatic failover across providers, a routing DSL, and model fusion. Because list price is passed through rather than marked up, a vendor price cut on any routed model is live the same day, and because failover is automatic, a provider having a bad afternoon does not become your outage. Qwen-Image-2.1 is not among the models we route and neither is any other Qwen-Image version — it is a local-only proposition — so this is not a pitch for the newcomer. It is the honest answer to "what do I use while I evaluate the newcomer."

The decision, stated plainly

Pick GPT Image 2 if the output is commercial, if you need an independently measured quality floor, if you want per-asset costs you can forecast and cap, or if you need the thing to work this week without a hardware purchase. Accept that you are renting, that you cannot fine-tune it, that you cannot run it offline, and that the per-image cost scales linearly with volume forever.

Pick Qwen-Image-2.1 if the output is research, evaluation or personal work, if you specifically want alpha that survives an edit loop rather than alpha that a parameter produced once, if you want to read the rewriting logic — the shipped PE-T2I and PE-I2I checkpoints are fine-tuned Qwen3.5-VL 9B models with a readable system_prompt.txt, which is more than you get from any hosted image API — or if you simply want to know what a 7B open-weights image model can do in September 2026 without paying per picture to find out.

Pick neither and route it if the deliverable is commercial and the deadline is real. That is not a cop-out; it is the option that does not require you to resolve a licence question you cannot yet get answered.