Qwen-Image-2.1 vs Ideogram 4.5: One You Rent, One You May Not Sell

Both Qwen-Image-2.1 and Ideogram 4.5 are on both Artificial Analysis image boards, and on both boards the released, downloadable Qwen-Image-2.1 is ahead. Text-to-image: 1,034 Elo at rank 18 against Ideogram 4.5 (High) at 1,014 and rank 34. Image editing: 1,074 at rank 18 against 1,064 at rank 23. That is a clean sweep on the scoreboard and a useless answer to the question most people actually have, because it compares a model you can download under a non-commercial research licence with a model you can rent by the image and ship commercially. Aliba​ba's Qwe​n team published Qwen-Image-2.1 with open weights on 20 September 2026; Ideogram 4.5 went live on 30 September 2026 as a hosted, closed model with a credit meter.

The other thing the Elo columns will not show you is that Ideogram's row rests on half the votes Qwe​n's does, and that Ideogram's own previous generation is the one shipping with weights. Both matter more than the twenty-point gap.

Half the sample, and what that does to the comparison

Look at the vote counts before the scores. On text-to-image, Qwen-Image-2.1's row is built from 5,286 blind comparisons and Ideogram 4.5 (High) from 2,247. On editing it is 5,536 against 2,382. Ideogram's rows are roughly 40% to 43% the size of Qwe​n's on the same boards, and part of that is simply age — the board has had twelve more days to accumulate votes on the Qwe​n model.

The confidence intervals make the same point. Ideogram 4.5 (High) carries a ±11 Elo interval on text-to-image and a ±11 on editing, against ±10 and ±9 for Qwen-Image-2.1. And on the editing board the raw rows are close enough to be worth stating per row:

• Text-to-image — Qwen-Image-2.1 at 1,034 (interval 1,024 to 1,044) versus Ideogram 4.5 (High) at 1,014 (interval 1,003 to 1,025). A twenty-point separation on intervals that just touch. Directional, not decisive.

• Image editing — Qwen-Image-2.1 at 1,074 (1,065 to 1,083) versus Ideogram 4.5 (High) at 1,064 (1,053 to 1,075). Ten points, intervals overlapping across roughly two Elo. This one is a coin-flip dressed as a ranking.

• Label — Qwen-Image-2.1 carries the Open Weights marker on both boards. Ideogram 4.5 (High) does not, on either.

The older Ideogram is the one with the weights

This is the detail that reframes the whole matchup. Ideogram announced that it plans to publish open weights for 4.5, and has not done so. What is downloadable from Ideogram today is the previous generation, and it is scored right alongside the new one:

• Ideogram 4.0 (Quality) — 1,011 Elo on text-to-image, rank 38, from 12,176 comparisons, with the Open Weights marker. That is 23 Elo below Qwen-Image-2.1 on more than twice the sample.

• Ideogram 4.0 — 1,004 Elo, rank 40, from 10,921 comparisons, Open Weights marker.

• Ideogram 4.0 Fast (Quality) and Ideogram 4.0 Fast — 984 and 968 Elo, both marked Open Weights.

• Ideogram 4.0 Instant — 982 Elo, rank 64, also Open Weights.

So the honest open-weights comparison is not Qwen-Image-2.1 against a closed hosted model. It is Qwen-Image-2.1 at 1,034 under a research-only licence against Ideogram 4.0 (Quality) at 1,011 under whatever licence those weights carry — a 23-point gap between the two strongest downloadable image models on the board, and a pair of licence questions rather than one. Ideogram 4.5 is the better model and it is the one you cannot download. That is the whole reason the two rows sit in the same table.

What the $0.22 buys

Ideogram 4.5's price is a ladder, not a number, and the board's price column picks the top rung. The service runs four rendering speeds — Turbo, Balanced, Quality and High — billed at $0.03, $0.06 and $0.22 per image for the Low, Medium and High tiers, on native 2K output at 2,048 pixels. The Artificial Analysis column records $220.0 per 1,000 images, which is the High tier and the same $0.22 stated per image. The board is quoting the most expensive configuration the model offers, not a typical bill.

Qwen-Image-2.1 has no price column at all — both of its board rows carry an explicit No API available note. You do not buy an image from it. You download the weight file, run it on hardware you pay for, and amortise that against however many images you generate. At high volume that is usually cheaper than $0.22 and it is always more work; at low volume it is always worse. Both of those are fine answers, and neither is a per-image price.

This is where the routing question lands for anyone who wants the hosted half without a second contract. OrcaRouter fronts 200+ models behind one OpenAI-compatible endpoint with provider list price passed through at zero markup, automatic failover across providers when one degrades, and a routing DSL that composes several models into a single call. Neither Qwen-Image-2.1 nor Ideogram 4.5 is on it — the image lines we serve are Google's Imagen tiers and Gemini image previews, xAI's Grok Imagine image endpoint and OpenAI's GPT-Image family. If the requirement is a callable, commercially licensed image model today, that list is the answer, and the scoreboard above is not the instrument that picks it.

The question the boards cannot answer

The licence is the axis with a settled answer, and it points the opposite way from the rankings. Qwen-Image-2.1 is released under the Qwe​n Research License Agreement dated 20 September 2026: non-commercial purposes only, with commercial use requiring a separate licence requested from the vendor. It is a change from the earlier Qwen-Image line, which shipped under Apache 2.0, and Aliba​ba's own Hugging Face repository reports the licence as other. The Open Weights marker on both board rows is accurate about downloadability and silent about this.

Ideogram 4.5, by contrast, is a commercial service. You rent the output and you own the right to use it — that is what the credit meter is for — and you never hold the weights. If the deliverable is a product with an image feature in it, that is a materially different position from "here is a file, and you may not use what comes out of it commercially."

There is a third path that neither column shows, and it is the one most teams end up on: run the downloadable model for internal work and evaluation, where the research licence is not in the way, and rent the hosted one for anything that reaches a customer. The boards give you a reason to prefer each, and the reason to prefer Qwen-Image-2.1 is a twenty-point Elo edge on a licence that stops at your own laptop. That is not a reason to put it in a product, and no scoreboard will ever tell you so.

What to watch is narrower than it looks. If Ideogram publishes the 4.5 weights it promised, the open-weights comparison against Qwen-Image-2.1 at 1,034 becomes a real fight rather than a 23-point proxy against Ideogram 4.0. Until then, the ranking and the licence point in opposite directions, and the licence is the one that survives the first contract review.